RWKV-7 Goose 0.1BRWKV-7 Goose 0.1B (approximately 190M parameters) is the smallest model in the RWKV-7 Goose series. Suitable for ultra-low-resource deployment with constant-memory inference. Uses RWKV-7 architecture with the Generalized Delta Rule. Trained on the World v2.8 corpus. Apache 2.0 licensed.
2025-03-18
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-6 Finch 1.6BRWKV-6 Finch 1.6B is the smallest model in the RWKV-6 Finch series, ideal for lightweight deployments requiring constant-memory inference. Apache 2.0 licensed.
2024-04-09
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-7 Goose 0.4BRWKV-7 Goose 0.4B (approximately 450M parameters) is a lightweight model from the RWKV-7 Goose series. Designed for edge deployment and resource-constrained environments where constant-memory O(1) inference is critical. Uses RWKV-7 architecture with the Generalized Delta Rule. Trained on the World v2.9 corpus. Apache 2.0 licensed.
2025-03-18
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-6 Finch 3BRWKV-6 Finch 3B is a mid-range model in the RWKV-6 Finch series, offering a balance between capability and deployment efficiency. Apache 2.0 licensed. Constant-memory inference with no KV cache.
2024-04-09
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-7 Goose 1.5BRWKV-7 Goose 1.5B is a mid-range model from the RWKV-7 Goose World3 series. Uses the seventh-generation RWKV architecture with Generalized Delta Rule for expressive dynamic state evolution. Trained on 3.1 trillion tokens from the multilingual World v3 corpus. Constant-memory inference with no KV cache. Apache 2.0 licensed.
2025-03-18
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-6 Finch 7BRWKV-6 Finch 7B is a flagship mid-size model from the RWKV-6 architecture series. Introduced alongside the Eagle and Finch paper (arXiv 2404.05892, April 2024). The Finch 14B model was subsequently derived by stacking two Finch 7B weights. Uses multi-headed matrix-valued states for improved language comprehension. Constant-memory inference. Apache 2.0 licensed.
2024-04-09
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-7 Goose 2.9BRWKV-7 Goose 2.9B is the largest released model in the RWKV-7 Goose World3 series. Built on the seventh-generation RWKV architecture with the Generalized Delta Rule and dynamic state evolution, it achieves competitive benchmark performance against transformer models of equivalent scale. Trained on 3.1 trillion tokens from the World v3 multilingual corpus (100+ languages, BF16). As a pure recurrent architecture, it requires constant O(1) memory during inference (no KV cache) and processes sequences in linear O(n) time. Licensed Apache 2.0.
2025-03-18
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
RWKV-6 Finch 14BRWKV-6 Finch 14B is the largest model in the RWKV-6 Finch series, created by stacking two 7B Finch models. Released September 3, 2024. Achieves strong performance on MMLU (56.05%), ARC, HellaSwag (57.69%), and Winogrande (74.43%). The RWKV-6 architecture uses matrix-valued states and dynamic data-driven recurrence, improving comprehension and in-context reasoning compared to RWKV-5 (Eagle). Constant-memory O(1) inference with no KV cache. Apache 2.0 licensed.
2024-09-03
Researched 177d ago
Infinite
No fixed token cap
Infinite context
No tracked provider route
LTM-2-miniLTM-2-mini is Magic's research prototype supporting a 100 million token context window, announced August 29, 2024. Uses a novel sequence-dimension algorithm approximately 1,000× more memory-efficient than transformer attention at this scale — requiring only a fraction of a single H100's HBM versus 638 H100s for Llama 3.1 405B at the same context length. Not publicly released for API access or self-hosting; Magic stated they were separately training a full LTM-2 model. Specialization: coding/software development. Source: https://magic.dev/blog/100m-token-context-windows
2024-08-29
Researched 183d ago
100M context
No tracked provider route
Llama 4 Scout 17B-16E InstructMeta's Llama 4 Scout is a 17-billion parameter mixture-of-experts model with 16 expert routing. Optimized for efficient inference on edge and cloud environments with strong multi-turn conversation capabilities. Available on Cloudflare Workers AI.
2025-04-05
Researched 164d ago
10M contextVisionMultimodalJSON
LTM-1LTM-1 (Long-Term Memory 1) is Magic's first model with a 5 million token context window, announced June 6, 2023. Designed to process entire codebases in context for AI-assisted software development. Architecture and parameter count not publicly disclosed. Not available as a public API; Magic used it in an early-access coding product. Source: https://magic.dev/blog/ltm-1
2023-06-06
Researched 183d ago
5M context
No tracked provider route
MiniMax-01MiniMax-01 combines MiniMax-Text-01 and MiniMax-VL-01, pairing a 456B-total-parameter MoE language model with multimodal understanding for long-context text generation and vision-language tasks.
2025-01-14
Researched 23d ago
4M contextVisionMultimodal
2024-08-27
Researched 183d ago
2M context
No tracked provider route
2024-08-01
Researched 183d ago
2M context
No tracked provider route
GPT-5.4 ProPremium extended-reasoning GPT-5.4 variant producing smarter and more precise responses. Replacement for o3-deep-research and o4-mini-deep-research. No prompt caching discount.
2026-03-01
Researched 32d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.6 LunaOpenAI's fast, low-cost GPT-5.6 model. Luna is the affordable tier in the Sol, Terra, and Luna lineup, optimized for latency-sensitive applications such as summarization, drafting, autocomplete, and routine automation. Generally available July 9, 2026 across ChatGPT, Codex, and the OpenAI API (model ID gpt-5.6-luna). Supports text and image inputs, reasoning, tool use, prompt caching, and Batch API with a 1,050,000-token context window and 128K max output.
2026-07-09
Researched 32d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.5 ProGPT-5.5 Pro is OpenAI's premium extra-compute deployment of GPT-5.5, released April 23, 2026. It uses the same underlying weights as GPT-5.5 standard with additional parallel test-time compute for harder tasks. Supports text and image inputs, reasoning effort control, tool use, structured outputs, code execution, a 1,050,000-token context window, and 128K max output. Reported benchmark scores: Terminal-Bench 2.1 78.2%, SWE-bench Pro 58.6%, GPQA Diamond 93.6%, ARC-AGI-2 high effort 83.3%, BrowseComp Pro compute 90.1%, and FrontierMath Tier 4 39.6%. Official pricing is $30/M input, $180/M output, $10/M batch input, and $45/M batch output; native cached input discount is not listed.
2026-04-23
Researched 145d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.6 TerraOpenAI's balanced mid-tier GPT-5.6 model. Terra delivers performance competitive with GPT-5.5 at about 2x lower cost, suited for high-volume production workloads such as customer support, document analysis, and internal tooling. Generally available July 9, 2026 across ChatGPT, Codex, and the OpenAI API (model ID gpt-5.6-terra). Supports text and image inputs, reasoning, tool use, prompt caching, and Batch API with a 1,050,000-token context window and 128K max output.
2026-07-09
Researched 132d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.5GPT-5.5 is OpenAI's fully retrained agentic model, released April 23, 2026. Optimised for agentic coding, computer use, knowledge work, and early scientific research. Achieves 82.7% on Terminal-Bench 2.0 (Codex CLI scaffold), 84.9% on GDPval, 58.6% on SWE-Bench Pro, 93.6% on GPQA Diamond, and 82.6% on SWE-Bench Verified (Vals.ai independent harness). Knowledge cutoff December 2025. Supports reasoning effort levels (none/low/medium/high/xhigh). Context window 1,050,000 tokens with a long-context surcharge above 272K tokens. Model ID: gpt-5.5.
2026-04-23
Researched 158d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.4GPT-5.4 is OpenAI's flagship frontier reasoning model, released March 5, 2026. It incorporates advances from GPT-5.3-Codex for coding and agentic workflows, and adds 'Thinking' mode with editable reasoning plans. Key capabilities include computer use (navigating interfaces via Playwright), image understanding and generation integration, full-stack web app generation, tool calling, and deep research. Knowledge cutoff is August 31, 2025. Model ID: gpt-5.4.
2026-03-05
Researched 32d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-5.6 SolOpenAI's flagship GPT-5.6 model and highest-capability tier in the Sol, Terra, and Luna naming system. GPT-5.6 Sol is built for demanding reasoning, long-horizon coding, agentic workflows, and cybersecurity tasks, introducing max reasoning effort and ultra multi-agent mode. Generally available July 9, 2026 across ChatGPT, Codex, and the OpenAI API (model IDs gpt-5.6-sol and alias gpt-5.6). Supports text and image inputs, reasoning, tool use, code execution, prompt caching, and Batch API with a 1,050,000-token context window and 128K max output.
2026-07-09
Researched 32d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6.1 TerraOpenAI GPT-6.1 Terra (API id gpt-6.1-terra), announced October 26, 2026, is the GPT-6.1 tier for fast, efficient visual reasoning, precise computer use, and clear, natural writing, positioned below GPT-6.1 Sol and GPT-6 Astra for complex coding and in-depth research. First-party docs: 1,050,000 context window (922,000 max input), 128,000 max output, knowledge cutoff Jun 1, 2026, text and image input to text output, reasoning.effort none/low/medium (default)/high/xhigh/max, function calling, structured outputs, prompt caching, Batch API, and computer use and code interpreter via Responses tools; fine-tuning not supported. Available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users.
2026-10-26
Researched 23d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6.2 AstraGPT-6.2 Astra (API id gpt-6.2-astra) is OpenAI's GPT-6.2 frontier model for demanding professional work, announced November 17, 2026 together with GPT-6.2 Sol. OpenAI positions it for complex coding, frontend and 3D work, computer use, and slides, spreadsheets, and documents that follow provided templates and field conventions. First-party docs list a 1,050,000-token context window, 922,000 max input tokens, 128,000 max output tokens, a May 18, 2026 knowledge cutoff, and text and image input with text output. It is a reasoning model (reasoning effort none is rejected) with function calling, structured outputs, prompt caching, Batch API support, and Responses API tools including web search, file search, code interpreter, hosted shell, apply patch, computer use, and MCP. OpenAI made it available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex and through the OpenAI API.
2026-11-17
Researched 1d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6.2 SolGPT-6.2 Sol (API id gpt-6.2-sol) is OpenAI's efficient GPT-6.2 model for everyday work, announced November 17, 2026 together with GPT-6.2 Astra. OpenAI describes it as aimed at multistep agentic workflows, coding, and clear writing. First-party docs list a 1,050,000-token context window, 922,000 max input tokens, 128,000 max output tokens, a June 1, 2026 knowledge cutoff, and text and image input with text output. It is a reasoning model (reasoning effort none is rejected) with function calling, structured outputs, prompt caching, Batch API support, and Responses API tools including web search, file search, code interpreter, hosted shell, apply patch, computer use, and MCP. OpenAI made it available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex and through the OpenAI API.
2026-11-17
Researched 1d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6.1 SolOpenAI GPT-6.1 Sol (API id gpt-6.1-sol), announced September 29, 2026, upgrades GPT-6 Sol toward near-Astra intelligence for complex coding, computer use, and professional work at one-fifth of Astra's standard input/output token prices. First-party docs: 1,050,000 context, 128,000 max output, knowledge cutoff Apr 30, 2026, text+image input to text output. Standard short-context list $2 input / $0.10 cached input / $2.50 cache writes / $10 output per 1M tokens (cached input 50% below GPT-6 Sol's $0.20). >272K input: 2x input/cache and 1.5x output. Batch/Flex 50% of Standard; Fast mode 2x. reasoning.effort low/medium/high/xhigh/max (none/minimal not supported). Function calling, structured outputs, prompt caching, computer use and code interpreter via Responses tools. Available ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu; not yet in Chat.
2026-09-29
Researched 23d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6 LunaOpenAI GPT-6 Luna, announced September 22, 2026 with GPT-6 Sol, is the most efficient GPT-6 tier for focused high-volume tasks below Sol and flagship Astra. First-party OpenAI API id gpt-6-luna. Docs: 1,050,000 context, 128,000 max output, knowledge cutoff May 18, 2026, text and image input to text output, reasoning effort none/low/medium/high/xhigh/max, function calling, structured outputs, prompt caching, Batch and Flex at 50% of Standard, Fast mode at 2x, and computer use and code interpreter via Responses tools. Standard short-context list is $0.10 input / $0.01 cached input / $0.125 cache writes / $0.50 output per 1M tokens (50% below GPT-5.6 Luna). Prompts with more than 272K input tokens use 2x input and cache rates and 1.5x output for the full request. Also generally available on Amazon Bedrock (openai.gpt-6-luna) and Microsoft Foundry. Available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu; Free and Go users can access Luna in the desktop app.
2026-09-22
Researched 56d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6 SolOpenAI GPT-6 Sol, announced September 22, 2026 with GPT-6 Luna, expands the GPT-6 family below flagship GPT-6 Astra for cost-efficient complex coding and agentic workflows. First-party OpenAI API id gpt-6-sol. Docs: 1,050,000 context, 128,000 max output, knowledge cutoff Apr 20, 2026, text and image input to text output, reasoning effort none/low/medium/high/xhigh/max, function calling, structured outputs, prompt caching, Batch and Flex at 50% of Standard, Fast mode at 2x, and computer use and code interpreter via Responses tools. Standard short-context list is $2 input / $0.20 cached input / $2.50 cache writes / $10 output per 1M tokens (50% below GPT-5.6 Sol promotional). Prompts with more than 272K input tokens use 2x input and cache rates and 1.5x output for the full request. Also generally available on Amazon Bedrock (openai.gpt-6-sol) and Microsoft Foundry. Available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu.
2026-09-22
Researched 32d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6 AstraGPT-6 Astra is OpenAI's flagship GPT-6 model, announced September 3, 2026 as a new generation of intelligence for computer use, browsing, software engineering, cybersecurity, science, and professional work. Official OpenAI API id gpt-6-astra. Docs list a 1,050,000-token context window, 922,000 max input tokens, 128,000 max output tokens, Apr 30, 2026 knowledge cutoff, image+text input to text output, reasoning.effort low/medium/high/xhigh/max (none is not supported), function calling, structured outputs, prompt caching, Batch API, computer use, and code interpreter. OpenAI API Standard list is $10 per 1M input tokens and $50 per 1M output tokens. Rolling out today to a limited set of organizations / Trusted Access Program; ChatGPT Plus, Pro, Business, and Enterprise, plus broader OpenAI API and AWS availability, in the coming days.
2026-09-03
Researched 23d ago
1.05M contextReasoningVisionMultimodalTool useFunctions
GPT-6 Luna DecisionsGPT-6 Luna Decisions is OpenAI's Decisions API surface (public beta) powered by GPT-6 Luna. Applications send shared text and/or image input plus typed questions (predicate / choice / score) to POST /v1/decisions and receive calibrated probabilities, choices with confidence, or rubric scores — about 10× faster than Luna via the Responses API per OpenAI docs. First-party model field is still gpt-6-luna (not a separate OpenAI model id). Billing is input tokens only at $0.10 per 1M tokens; no cache-read, cache-write, or output-token charges. Regional processing premiums and long-context input multipliers apply. Vercel AI Gateway exposes the same surface as openai/gpt-6-luna-decisions. Belongs on /decision-models alongside d1 and other decision models. Proprietary; no public weights.
2026-10-06
Researched 42d ago
1.05M contextReasoningVisionMultimodalJSON
Arrow 2 TelosArrow 2 Telos (API model id arrow-2-telos) is QuiverAI's most capable generally available Arrow model for demanding vector work and longer contexts. Launch blog: combines Arrow speed with frontier-model refinement for complex styles, compositions, and requirements; choose Telos for harder briefs. Docs: inputs text/image/SVG; output SVG; context 1,050,000 tokens; max output 65,536; SVG generation + image vectorization supported; editing/animation and Responses when catalog-enabled; token_usage billing at 50% higher rates than Arrow 2 in every token category. Proprietary hosted — no public weights.
2026-09-07
Researched 62d ago
1.05M contextVisionMultimodalTool useFunctionsPrompt cache
MiMo-V2.5Xiaomi MiMo-V2.5 is the lower-cost native omnimodal sibling in the MiMo-V2.5 series. OpenRouter describes it as supporting text, image, audio, and video inputs with text output, Pro-level agentic performance at roughly half the inference cost, and improved multimodal perception over MiMo-V2-Omni. Xiaomi's official April 22 release page highlights MiMo-V2.5 alongside MiMo-V2.5-Pro in benchmark data and says the V2.5 series will be open-sourced soon; no public weights/license were verified at research time.
2026-04-22
Researched 29d ago
1M contextReasoningVisionMultimodalTool useFunctions
MiMo-V2.5-ProXiaomi's April 22, 2026 public-beta flagship in the MiMo-V2.5 series. The official Xiaomi MiMo page describes MiMo-V2.5-Pro as its most capable model to date, focused on general agentic capability, complex software engineering, long-horizon tasks, and ultra-long-context instruction following. OpenRouter lists it as text-to-text with 1,048,576 token context, 131,072 max completion tokens, reasoning controls, tool use, and response_format support. Xiaomi says the V2.5 series will be open-sourced soon, but no public weights/license were verified at research time.
2026-04-22
Researched 29d ago
1M contextTool useFunctionsJSONPrompt cache
MiMo-V2.6-FlashXiaomi MiMo-V2.6-Flash is Xiaomi MiMo's September 21, 2026 efficiency-balanced open-weights MoE (309B total / 15B active), sibling of MiMo-V2.6-Pro. First-party HF checkpoint XiaomiMiMo/MiMo-V2.6-Flash-RL (MIT; ungated; 1M context) is the released Flash RL weights; product/API id is mimo-v2.6-flash. First-party news: the series comprises two native fully multimodal models (Pro and Flash); Flash comprehensively outperformed MiMo-V2.5-Pro. Product card: full-modality high-intelligence low-cost reasoning for high-frequency and large-scale professional workflows; context 1M; max output 128K; input text, image, video, and audio. Overseas API pricing (pay-as-you-go): $0.0028 cache-hit / $0.14 cache-miss input and $0.28 output per 1M tokens; batch $0.0014 / $0.07 / $0.14. Sparse MoE MiMoV2ForCausalLM; 48 layers (39 SWA / 9 GA); 256 routed experts / 8 activated per token; hybrid SWA+GA backbone; native omnimodal (text/image/video/audio); 681M MiMo ViT vision encoder; 5-layer SWA MTP/DFlash speculative decoder.
2026-09-21
Researched 29d ago
1M contextReasoningVisionMultimodalAudioTool use
MiMo-V2.6-ProXiaomi MiMo-V2.6-Pro is Xiaomi MiMo's September 21, 2026 flagship open-weights MoE (1.02T total / 42B active), succeeding proprietary MiMo-V2.5-Pro. First-party HF checkpoint XiaomiMiMo/MiMo-V2.6-Pro-RL (MIT; ungated; 1M context) is the released Pro RL weights; product/API id is mimo-v2.6-pro. HF README: native omnimodal + long-horizon agents; You Only RL Once mixed RL across coding/agents/visual/cybersecurity. Official product page (mimo.mi.com) and @XiaomiMiMo announce open-sourcing of Pro and Flash with tech report. Overseas API pricing matches V2.5-Pro: $0.0036 cache-hit / $0.435 cache-miss input and $0.87 output per 1M tokens (pay-as-you-go). Sparse MoE MiMoV2ForCausalLM; 70 layers (60 SWA / 10 GA); 384 routed experts / 8 activated per token; hybrid SWA+GA backbone; native omnimodal (text/image/video/audio); 681M MiMo ViT vision encoder; MTP blocks.
2026-09-21
Researched 29d ago
1M contextReasoningVisionMultimodalAudioTool use
Kimi K3Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship multimodal model for long-horizon coding, knowledge work, deep reasoning, and agentic workflows. It uses Kimi Delta Attention, Attention Residuals, and a sparse MoE design (16 of 896 experts active), supports a 1,048,576-token context window, text/image/video input, always-on reasoning, ToolCalls, strict JSON Schema structured output, automatic context caching, and partial mode through Moonshot's OpenAI-compatible API.
2026-07-14
Researched 33d ago
1M contextReasoningVisionMultimodalTool useFunctions
Kimi K3.1 ProKimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, a post-trained version of Kimi K3. It has a 1,048,576-token context window, always-on reasoning with low, high and max effort, image and video input, tool calling and strict JSON Schema structured output. Released 2026-10-23 on Kimi.ai, Kimi Work, Kimi Code and the Kimi API; Moonshot published the weights on Hugging Face under the Kimi K3.1 Pro License.
2026-10-23
Researched 14d ago
1M contextReasoningVisionMultimodalTool useFunctions
Nemotron 3 Super-120B-A12BNVIDIA Nemotron 3 Super-120B-A12B is a 120B total / 12B active hybrid Latent MoE model with interleaved Mamba-2 and MoE layers for agentic, reasoning, and conversational tasks. Fireworks lists the NVFP4 variant for on-demand deployment with 262k context.
2026-03-11
Researched 23d ago
1M contextReasoningFunctionsJSONPrompt cache
Hy4 flashHy4 flash is Tencent Hy’s multimodal mixture-of-experts model for everyday coding, agents and document understanding. It accepts text and images with a 1,048,576-token context. Its architecture combines sparse and sliding-window attention, LatentMoE and n-gram embeddings. Tencent publishes instruct, FP8 and pre-trained base weights under the custom Tencent Hy License, which includes additional commercial terms for large MaaS businesses.
2026-11-16
Researched today
1M contextReasoningVisionMultimodalTool useFunctions
Gemini 4 HydrogenGemini 4 Hydrogen is Google DeepMind's fast, lower-cost Gemini 4 model, released generally available on November 12, 2026 in the Gemini API, Google AI Studio and Vertex AI (Gemini Enterprise Agent Platform). Google positions it for everyday coding, agentic workflows and multimodal work, alongside Gemini 4 Carbon for the hardest problems; its model card says it is adapted from Gemini 4 Carbon. It takes text, image, video, audio and PDF input and returns text, with a unified 1,048,576-token context window and long decode up to 1,048,576 output tokens (65,536 by default). It adds a fourth xhigh thinking level alongside low, medium and high, and supports function calling, structured outputs, code execution, computer use, Google Search and Maps grounding and implicit caching. Paid-tier pricing is $1.50 per 1M input tokens and $7.50 per 1M output tokens, with a free tier in AI Studio. Knowledge cutoff is March 2026. Hosted only; weights are not distributed.
2026-11-12
Researched 5d ago
1M contextReasoningVisionMultimodalAudioFunctions
Muse EmberMuse Ember is Meta Superintelligence Labs' smallest and most efficient Muse model, introduced November 9, 2026, for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta says it is the first model trained using Muse Wonder Full. It takes text, image, audio and video input and returns text, with a 1M-token context window and up to 512K output tokens, and supports the Responses API. It is available on Muse Code and the Meta Model API as muse-ember-1.0 at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens.
2026-11-09
Researched 8d ago
1M contextReasoningVisionMultimodalAudioTool use
Particle-1Particle-1 is Particle AI's model for agentic coding and software engineering, post-trained on DeepSeek V4.1 Flash to finish tasks with fewer tokens. It takes text and images, supports reasoning and tool use, and has a 1,048,576-token context window with up to 131,072 output tokens. Available on the Particle API, OpenRouter and Vercel AI Gateway.
2026-10-28
Researched 20d ago
1M contextReasoningVisionMultimodalTool useFunctions
GLM-LegalGLM-Legal is Z.ai's legal-domain model, built on GLM-5.3-Flash with legal mid-training and reinforcement-learning post-training across legal agent environments. Z.ai reports a 19.2% all-pass rate on the Harvey Legal Agent Benchmark, 12.5 points above its base model, as evaluated by Vals.
2026-10-12
Researched 16d ago
1M contextReasoningVisionMultimodalTool useFunctions
Ripple-65B PreviewRipple-65B Preview is Ricursive Intelligence's open-weight 65B-parameter dense Transformer, released November 4, 2026 under Apache 2.0. It is a general-purpose variant of the in-house model behind Ricursive's AI chip design platform, pretrained on 12 trillion tokens with support for up to 1M tokens of context. Weights are on Hugging Face.
2026-11-04
Researched 13d ago
1M contextTool useFunctionsPrompt cache
Kimi K3.1 Pro HighspeedKimi K3.1 Pro Highspeed is the high-speed variant of Moonshot AI's Kimi K3.1 Pro, offered on the Kimi API with the same 1,048,576-token context window. It keeps always-on reasoning with low, high and max effort, image and video input, tool calling and strict JSON Schema structured output, at double the K3.1 Pro token price. Released 2026-10-23 alongside K3.1 Pro.
2026-10-23
Researched 14d ago
1M contextReasoningVisionMultimodalTool useFunctions
AmpersandAmpersand (v0.1, API id ampersand-v0.1) is humans&'s collaborator model, built on GLM 5.3 and trained with RL to understand the long-term consequences of its actions and help groups of people reach shared goals: reasoning together, asking questions, surfacing context and working through disagreements. Text in/out, 1M-token context, API-only.
2026-10-20
Researched 27d ago
1M contextReasoningTool useFunctionsPrompt cache
Hy4 flash BaseHy4 flash Base is Tencent’s pre-trained multimodal checkpoint in the Hy4 flash line, distributed alongside its post-trained instruct model. It accepts text and image inputs and uses hybrid attention, LatentMoE and n-gram embeddings. The weights are provided under the Tencent Hy License with additional MaaS commercial terms.
2026-11-18
Researched today
1M contextVisionMultimodal
No tracked provider route
Spark X2.6 1.7BSpark X2.6 1.7B is iFlytek's (SparkLLM / XHToken) second Spark-X2.6 model, a compact open-weight language model for on-device agents. It keeps the Spark-X2.5-1.7B architecture and improves coding, agentic tool use, long-context, multilingual and reasoning results through better post-training, offers three thinking levels (none, low, high), and can be served with up to 1,048,576 tokens of context. Weights (plus a separate base checkpoint) are on Hugging Face under Apache 2.0.
2026-10-27
Researched 22d ago
1M contextReasoningTool useFunctions
No tracked provider route
Spark X2.6 4BSpark X2.6 4B is iFlytek's compact open-weight language model for on-device agents, announced by SparkLLM on October 20, 2026. It keeps the Spark-X2.5-4B architecture and improves coding, agentic tool use, long-context, multilingual and reasoning results through better post-training, offers three thinking levels (none, low, high), and can be served with up to 1,048,576 tokens of context. Weights (plus a separate base checkpoint) are on Hugging Face under Apache 2.0.
2026-10-20
Researched 28d ago
1M contextReasoningTool useFunctions
No tracked provider route
2026-03-19
Researched 188d ago
1M context
No tracked provider route
Claude Opus 5.5Anthropic's Claude Opus 5.5, first model in the Claude 5.5 family, released September 22, 2026. First-party positioning: long-running agentic coding and knowledge work; performs at the level of Claude Fable 5.1 on most work; typical workloads about 40% less cost than Opus 5 at default settings (list input/output 20% below Opus 5; cache reads 60% below). Claude API id claude-opus-5-5. List price $4 per 1M input and $20 per 1M output; cache reads $0.20 per 1M tokens (0.05x base input); 5-minute cache writes $5; 1-hour cache writes $8; Batch API 50% off ($2/$10). Fast mode (Claude API research preview) is $8/$40. 1M-token context, 128K max output (Batch API beta up to 300K with output-300k-2026-03-24), adaptive thinking always on (default effort medium), text and image input to text output, vision, and tool use. Available on Claude for Pro, Max, Team, and Enterprise and on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Reliable knowledge cutoff June 2026.
2026-09-22
Researched 33d ago
1M contextReasoningVisionMultimodalTool useFunctions
Inkling 2 SmallInkling 2 Small is Thinking Machines Lab's first second-generation Inkling model, released November 5, 2026. It is a 594B-total / 17B-active mixture-of-experts model for text and image inputs with text output, a context window of up to 1M tokens and controllable reasoning effort. It launched for inference and fine-tuning on Tinker; on November 12, 2026 Thinking Machines released its weights on Hugging Face under Apache 2.0 (NVFP4, plus an MXFP8 variant), and it is also served by Together AI, Cloudflare Workers AI, DeepInfra and Vercel AI Gateway at $0.30 per 1M input and $1.00 per 1M output tokens.
2026-11-05
Researched 5d ago
1M contextReasoningVisionMultimodalTool useFunctions
Gemini 3.5 FlashGemini 3.5 Flash is Google DeepMind's generally available Flash model for sustained frontier-level performance on agentic and coding tasks. It supports multimodal inputs, native thinking, tool and function calling, structured outputs, code execution, search grounding, batch processing, and long contexts up to 1M tokens.
2026-05-19
Researched 159d ago
1M contextReasoningVisionMultimodalAudioTool use
Gemini 3.1 Flash-LiteGemini 3.1 Flash-Lite is Google's generally available low-latency Gemini 3.1 model, launched May 7, 2026. It is optimized for high-volume, cost-sensitive workloads with text, image, and video inputs, a 1M token context window, and a 66K token maximum output. The GA model uses the stable API ID gemini-3.1-flash-lite and replaces gemini-3.1-flash-lite-preview, which is scheduled to shut down on May 25, 2026. Pricing is $0.25 per 1M input tokens and $1.50 per 1M output tokens.
2026-05-07
Researched 32d ago
1M contextVisionMultimodalTool useFunctionsJSON
Ling-3.2-flashLing-3.2-flash is Ant Ling's (InclusionAI) native multimodal MoE, launched November 5, 2026 with 560B total / 25B active parameters and a 1M-token context window. It natively accepts image, text, and video input for complex visual understanding, multimodal long-horizon agents, 3D understanding, and medical multimodal tasks. It is available via Ant Ling's OpenAI- and Anthropic-compatible APIs.
2026-11-05
Researched 12d ago
1M contextVisionMultimodalTool useFunctions
No tracked provider route
Nex-N2.5-MaxNex-N2.5-Max is Nex AGI's trillion-scale text-only MoE agentic model in the Nex-N2.5 family, released September 8, 2026. Apache-2 weights on Hugging Face (nex-agi/Nex-N2.5-Max). 1.6T total / 49B active MoE with a native 1,048,576-token context window (config max_position_embeddings). Supports reasoning, function calling, and tool use; no vision or multimodal inputs.
2026-09-08
Researched 70d ago
1M contextReasoningTool useFunctions
No tracked provider route
Claude Haiku 5.5Anthropic's Claude Haiku 5.5, third model in the Claude 5.5 family, released October 7, 2026. First-party positioning: cheapest, fastest, and most capable small model yet; built for high-volume, cost-sensitive work (summaries, compaction, classification, routing, subagents, live support, browser/computer use). First Haiku with effort controls and adaptive thinking (default effort medium). Claude API id claude-haiku-5-5. Prompt-length pricing: ≤100K tokens $0.10/$0.50 per 1M input/output; >100K $0.50/$2.50. Cache (≤100K): 5m write $0.125, 1h write $0.20, read $0.01; (>100K): $0.625/$1/$0.05. Batch API 50% off. 1M-token context, 128K max output (Batch API beta up to 300K), text and image input to text output, vision, tool use, prompt caching. Available on Claude apps (Free/Pro/Max/Team/Enterprise), Claude Code, Claude API, Amazon Bedrock (incl. GovCloud US), Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Reliable knowledge cutoff June 2026; retirement not sooner than October 7, 2027. Anthropic claims ~75% lower run cost vs Haiku 4.5 for most tasks.
2026-10-07
Researched 42d ago
1M contextReasoningVisionMultimodalTool useFunctions
Claude Sonnet 5.5Anthropic's Claude Sonnet 5.5, second model in the Claude 5.5 family, released September 28, 2026. First-party positioning: faster, lower-cost complement to Claude Opus 5.5; strongest at well-scoped everyday tasks, bug fixes, and polished documents/slides/spreadsheets; clear upgrade over Claude Sonnet 5 (30%+ faster output; up to 30% less cost per task via fewer tokens — list price unchanged vs Sonnet 5). Claude API id claude-sonnet-5-5. List price $2 per 1M input and $10 per 1M output; cache reads $0.10 / MTok; 5m cache writes $2.50; 1h cache writes $4; Batch API 50% off ($1/$5). 1M-token context, 128K max output (Batch API beta up to 300K), adaptive thinking (default effort high; between_tools turns off up-front thinking), text and image input to text output, vision, tool use. Available on Claude apps and on Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Reliable knowledge cutoff June 2026.
2026-09-28
Researched 23d ago
1M contextReasoningVisionMultimodalTool useFunctions
Gemini 4 CarbonGemini 4 Carbon is a Google DeepMind reasoning model that Google announced on October 28, 2026 for a subset of cyber defenders and trusted partners in its Fairwind Program, and made generally available on November 10, 2026 through the Gemini API, Gemini Enterprise Agent Platform (Vertex AI), Google Antigravity, the Gemini app and AI Mode in Search for Google AI Ultra subscribers. Google describes it as its most capable reasoning model, built for sustained agentic coding, multimodal knowledge work, scientific reasoning and the cybersecurity defense lifecycle. It takes text, image, video, audio and PDF input and returns text. Input and output share a 1,048,576-token limit; one synchronous request returns up to 65,536 output tokens, and long decode extends output to 1,048,576 tokens. Carbon belongs to the Gemini 4 family, builds on Gemini 4 Argon and launched at Argon's introductory price.
2026-10-28
Researched 8d ago
1M contextReasoningVisionMultimodalAudioFunctions
Muse Spark OpenMuse Spark Open is the open-weight release of Meta Superintelligence Labs' Muse Spark, an 816-billion-parameter mixture-of-experts model that activates about 42 billion parameters per token. Released October 15, 2026 under the Muse Model License, it is built for coding and agentic work, takes text, image, video and audio input, returns text, and has a 1M-token context window. Weights are on Hugging Face in BF16 and NVFP4.
2026-10-15
Researched 34d ago
1M contextReasoningVisionMultimodalAudioTool use
Gemini 4 ArgonGemini 4 Argon is a Google DeepMind reasoning model, announced on September 30, 2026 for trusted cyber defenders and testers in Google's Fairwind Program and made generally available on October 13, 2026 through the Gemini API, Gemini Enterprise Agent Platform, Google Antigravity and the Gemini app. Google built it for long-horizon software engineering, legal and finance work, cybersecurity defense and science. It takes text, image, video, audio and PDF input and returns text. Input and output share a 1,048,576-token limit; one synchronous request returns up to 65,536 output tokens, and long decodes extend output up to 1,048,576 tokens. It supports low, medium and high thinking levels, function calling, structured outputs, code execution, search grounding and implicit caching.
2026-09-30
Researched 32d ago
1M contextReasoningVisionMultimodalAudioFunctions
Muse WonderMuse Wonder is a Meta Superintelligence Labs model released October 12, 2026, for agentic coding, long-running agents and computer use. Meta says it builds on a larger base model than the Muse Spark models. It takes text, image, video and audio input and returns text, with a 1M-token context window. A larger sibling, Muse Wonder Full, is offered only through a Muse Code subscription.
2026-10-12
Researched 32d ago
1M contextReasoningVisionMultimodalAudioTool use
Mistral Large 4Mistral Large 4 (API id mistral-large-4-0, public preview launched 6 October 2026) is Mistral AI's multimodal mixture-of-experts model: 1.05T total parameters, 49B active per token (52B including embeddings and output layers), and a 1.6B vision encoder, with a 1M-token context window. On 29 October 2026 Mistral released the weights of an improved version on Hugging Face under the Mistral Ecosystem Open-Weight License (a modified MIT license that requires a separate commercial license to run a Model-as-a-Service business).
2026-10-06
Researched 20d ago
1M contextVisionMultimodalTool useFunctionsJSON
Nemotron 3.5 LightningNVIDIA Nemotron 3.5 Lightning (HF nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B) is an open-weight hybrid Mamba-Transformer mixture-of-experts text model with 30B total and 3B active parameters (NemotronHForCausalLM), released 11 August 2026. NVIDIA lists context up to 1M tokens (256K used for single-H100 deployment). Weights ship in BF16, NVFP4 and speculative-decoding (DSpark/DFlash) variants plus a base checkpoint under OpenMDW 1.1. It is offered on build.nvidia.com and OpenRouter (free tier).
2026-08-11
Researched 44d ago
1M contextReasoningTool useFunctionsPrompt cache
Muse Spark 1.2 ContributorMuse Spark 1.2 Contributor is the Contributor-tier Model API id (muse-spark-1.2-contributor) for Muse Spark 1.2. Distinct catalog ID (same pattern as muse-spark-1.3-contributor). 1,048,576-token context; multimodal inputs. Contributor mode: discounted tokens in exchange for permission to use content to train future Meta models. FP pricing: $0.10/$0.20 per 1M in/out (cached input $0.002).
1M contextReasoningVisionMultimodalAudioTool use
Pareto 26.10Unbiased AI blended/composite model Pareto 26.10. OpenAI+Anthropic compatible API. Stable GA Oct 15, 2026 (preview Oct 1). FP pay-as-you-go $0.80/$3.20.
2026-10-15
Researched 20d ago
1M contextMultimodalPrompt cache
BeamBeam is an open-weight sparse mixture-of-experts language model with 501B total and 23B active parameters, built for coding, reasoning, and long-horizon agentic work and trained from scratch. It is text-only, with a 1M-token context window, up to 128K output tokens, five reasoning effort levels (low to max, default medium), tool calling, and structured outputs, and a June 30, 2026 knowledge cutoff. Reflection announced Beam on 5 October 2026 as a waitlisted API preview and released the weights on Hugging Face under Apache 2.0 on 16 October 2026, in BF16, FP8, and NVFP4 checkpoints, alongside self-serve access through the Reflection API (model id Beam-1.0-501B-A23B).
2026-10-05
Researched 33d ago
1M contextReasoningTool useFunctionsJSONPrompt cache
Gemini 3.8 FlashGemini 3.8 Flash is Google DeepMind's generally available most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows, released September 2, 2026. It accepts text, image, video, audio, and PDF inputs and returns text, with a 1,048,576-token context window, up to 65,536 output tokens, and Gemini API support for thinking (low/medium/high; minimal is not supported), function calling, tool use, structured outputs, code execution, prompt caching, search grounding, URL context, computer use (preview), and batch, flex, and priority consumption. Official model ID: gemini-3.8-flash. Sibling of live gemini-3.7-flash under a new family gemini-3.8.
2026-09-02
Researched 77d ago
1M contextReasoningVisionMultimodalAudioTool use
Muse Spark 1.3Muse Spark 1.3 is Meta Superintelligence Labs' latest Muse Spark checkpoint, released September 2, 2026. Official Meta Model API model ID: muse-spark-1.3. First-party models docs: the latest version, recommended for new work, tuned for agentic workflows (multi-step tool, browser, and long-horizon tasks) with improved coding over 1.2; default model in docs examples. Product line (dev.meta.ai): Built for long-horizon coding and agentic work. Cleaner code, fewer tokens. First-party research post: rolling out today in Muse Code and Meta Model API; previously available reasoning modes available today with max reasoning coming shortly after additional safety testing. Official table: text, image, video, audio*, PDF input and text output; 1,048,576-token context. Official quickstart limits: context 1048576, output 131072; input modalities text/image/pdf/video, output text. Audio understanding in 1.3 is currently not fully supported (use 1.2 or Muse Voice Transcribe for audio). Tool/function calling, structured outputs, prompt caching, search grounding, and reasoning_effort (minimal/low/medium/high/xhigh; none is not supported).
2026-09-02
Researched 44d ago
1M contextReasoningVisionMultimodalTool useFunctions
Muse Spark 1.3 ContributorMuse Spark 1.3 Contributor is the Contributor-tier catalog ID (muse-spark-1.3-contributor) of Muse Spark 1.3 on Meta Model API. Official models page lists it as a distinct Model ID on the Contributor tier with the same 1,048,576-token context and the same input modalities as muse-spark-1.3 (text, image, video, audio*, PDF) and text output. First-party pricing page lists it with its own Contributor rates. Heavily discounted token pricing in exchange for permission to use prompts and completions to train future Meta models. Audio understanding in 1.3 is currently not fully supported. Not a separately published checkpoint or weights drop.
2026-09-02
Researched 44d ago
1M contextReasoningVisionMultimodalTool useFunctions
Hunyuan Hy4 PreviewTencent Hunyuan Hy4 preview is a Mixture-of-Experts flagship language model released and open-sourced August 28, 2026. First-party Tencent release: 770B total parameters with 49B activated per token, context window exceeding 1M tokens, aimed at coding, office productivity, game development, and scientific research. Hugging Face model card (tencent/Hy4-preview) documents a 78-layer MoE backbone (256 routed experts, top-8 plus 1 shared expert), Gated DeepSeek Sparse Attention with IndexCache, iHC residual streams, native MTP speculative decoding, Apache License 2.0 weights (also Hy4-preview-FP8), and OpenAI-compatible serving via vLLM/SGLang with tool-call and reasoning parsers. Public access: Hugging Face/ModelScope/GitCode/CNB weights; API via Tencent Cloud TokenHub and OpenRouter (model id tencent/hy4-preview); also WorkBuddy, CodeBuddy, Yuanbao, and ima. Official API list pricing USD 0.834 / 2.501 per 1M input/output tokens and USD 0.042 per 1M cache hits.
2026-08-28
Researched 79d ago
1M contextReasoningTool useFunctionsPrompt cacheFine-tune
Gemini 3.7 FlashGemini 3.7 Flash is Google DeepMind's generally available multimodal workhorse model for coding and agents, released August 13, 2026. It accepts text, image, video, audio, and PDF inputs and returns text, with a 1,048,576-token context window, up to 65,536 output tokens, and Gemini API support for thinking (low/medium/high; minimal is not supported), function calling, tool use, structured outputs, code execution, prompt caching, and batch processing. Official model ID: gemini-3.7-flash.
2026-08-13
Researched 86d ago
1M contextReasoningVisionMultimodalAudioTool use
Muse Spark 1.2Muse Spark 1.2 is Meta Superintelligence Labs' proprietary multimodal reasoning model for agentic coding and long-context work, released August 5, 2026. It supports a 1,048,576-token context window and up to 131,072 output tokens through the Meta Model API. No public weights release at launch.
2026-08-05
Researched 32d ago
1M contextReasoningVisionMultimodalTool useFunctions
Gemini 3.5 Flash-LiteGemini 3.5 Flash-Lite is Google DeepMind's generally available multimodal model in the Gemini 3.5 family. It accepts text, image, video, audio, and PDF inputs and returns text, with a 1,048,576-token context window, up to 65,536 output tokens, and Gemini API support for code execution, prompt caching, and batch processing. In Google AI Studio, select Gemini 3.5 Flash-Lite when lower token cost and high throughput are the priority; select Gemini 3.6 Flash when higher capability matters more.
2026-07-21
Researched 120d ago
1M contextReasoningVisionMultimodalAudioTool use
Gemini 3.6 FlashGemini 3.6 Flash is Google DeepMind's generally available multimodal model. It accepts text, image, video, audio, and PDF inputs and returns text, with a 1,048,576-token context window, up to 65,536 output tokens, and Gemini API support for code execution, prompt caching, and batch processing. In Google AI Studio, select Gemini 3.6 Flash when higher capability is the priority; select Gemini 3.5 Flash-Lite when lower token cost and high throughput matter more.
2026-07-21
Researched 32d ago
1M contextReasoningVisionMultimodalAudioTool use
Nemotron 3.5 Super VLNemotron 3.5 Super VL is NVIDIA's multimodal Super model in the Nemotron 3.5 family, a 120-billion-parameter mixture-of-experts model with about 12 billion active parameters. It adds vision to Super's hybrid Mamba-Transformer MoE, with a 1M-token context window, and NVIDIA reports it can read text in images, video, web pages and PDFs. Available as open weights and through hosted APIs.
2026-10-15
Researched 34d ago
1M contextReasoningVisionMultimodalTool useFunctions
Gemini Omni 1.1 FlashGemini Omni 1.1 Flash is Google DeepMind's generally available video generation and editing model, released August 27, 2026. It provides fast conversational video generation and editing through the Gemini API and Interactions API, with scene extension, interpolation, resolution upscaling, and support for video references. It accepts text, image, and video input and outputs video with native audio at 360p, 720p, 1080p, or 4K (3s–10s at 24 FPS). Official model ID: gemini-omni-1.1-flash.
2026-08-27
Researched 82d ago
1M contextVisionMultimodalAudio
Doubao Seed EvolvingDoubao Seed Evolving is ByteDance's rapidly iterating coding and agent model on Volcengine Ark. It is listed as a rapidly iterating model (快速迭代模型) and supports deep thinking, text generation, multimodal understanding, GUI tasks, tool use, and structured output. The documented context window and maximum input are 1024K tokens; maximum output and thinking-chain length are 256K tokens.
1M contextReasoningVisionMultimodalTool useFunctions
Muse Wonder FullMuse Wonder Full is a Meta Superintelligence Labs model released October 12, 2026, and the larger sibling of Muse Wonder, for agentic coding, long-running agents and computer use. It takes text, image, video and audio input and returns text, with a 1M-token context window. It is available only through a Muse Code subscription, not the Meta Model API.
2026-10-12
Researched 37d ago
1M contextReasoningVisionMultimodalAudioTool use
No tracked provider route
Kolibri 1Kolibri (Kolibri 1, HF Aleph-Alpha/Kolibri-1) is Aleph Alpha's open-weight English–German mixture-of-experts reasoning model, released 3 October 2026 and trained from scratch in Germany. It has 78B total parameters (78,103,074,560) with 3.46B active per token: 50 layers, 384 routed experts plus 1 shared, top-6 sigmoid routing, and 4:1 sliding-window-to-full GQA attention. Native context is 262,144 tokens. Because RoPE is applied only in the sliding-window layers, Aleph Alpha validates up to 1,048,576 tokens via a max_position_embeddings override, and recommends at most 262,144 for serving efficiency and complex tasks. It is text-in, text-out with controllable reasoning effort (low/medium/high/none) and Hermes-style tool calling. Knowledge cutoff is 18 June 2026. It was pre-trained on 20T bilingual tokens (~62.5% EN, ~23.9% DE, ~13.6% code), then mid-training (3.44T), long-context extension (201B), SFT and RL. Weights are released under Apache 2.0 as FP8 (Kolibri-1, ~78 GB, 2× H100 or 1× H200 minimum) and BF16 (Kolibri-1-BF16). Serving uses vLLM with the aleph-alpha-inference plugin. No hosted per-token API price is published; Aleph Alpha offers enterprise deployment via sales. It is a new family distinct from Aleph Alpha's earlier proprietary Luminous line. Kolibri Origin (30B-A3B) was an internal, unreleased predecessor.
2026-10-03
Researched 44d ago
1M contextReasoningTool useFunctionsJSON
No tracked provider route
Spark X2.5 1.7BSpark-X2.5-1.7B is the smaller SparkLLM / XHToken (iFlytek) post-trained on-device general language model in the Spark-X2.5 series. Hugging Face card XHToken/Spark-X2.5-1.7B (Apache License 2.0 LICENSE blob, Copyright 2026 XHToken) documents Spark2_5ForCausalLM with the same hybrid 3:1 sliding-window / full-attention pattern, native 1,048,576-token context, safetensors 1,707,657,216 parameters (card title 1.7B; HF size widget rounds to 2B), thinking mode, tool-call parser spark25, and Llama-Factory fine-tuning. Same first-party GitHub README as the 4B: ~20T pretrain, SFT plus RL / MOPD from base XHToken/Spark-X2.5-1.7B-Base; 200+ languages; agent/coding/reasoning on-device. Serving: SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio. Xingchen MaaS listed on SparkLLM DEV as limited-time free; first-party Token Plan docs on 2026-09-01 list Spark-X2 / X2-Agent / X2-Flash only, not X2.5 IDs or USD rates. Sibling Spark-X2.5-4B.
2026-08-31
Researched 78d ago
1M contextReasoningTool useFunctionsFine-tune
No tracked provider route