Reasoning

Lightweight reasoning variant of Microsoft Phi-4 Mini optimized for fast inference.

Pricing not tracked / 1M tokens

1 route

Magistral Small 2506 is MistralAI's Magistral model focused on step-by-step reasoning. It offers a 128K-token context window.

Pricing not tracked / 1M tokens

1 route

Tencent HunYuan's Hy3 Preview is a high-efficiency Mixture-of-Experts language model for agentic and production workflows. OpenRouter lists tencent/hy3-preview as released Apr 22, 2026 with 262,144 context, preview pricing, reasoning controls, and tool-use support. Hugging Face Transformers documents Hy3-preview as a Tencent HunYuan MoE model with a dense-MoE hybrid architecture, 192 routed experts, and one always-active shared expert per MoE layer. SCMP's Apr 23 coverage reports Tencent described HY3-Preview as a new flagship model developed by the HunYuan and Yuanbao teams with 295B parameters. Treat release metadata as high confidence for existence/context and medium confidence for exact parameter count until Tencent publishes a primary technical card.

$0.066 in / $0.260 out / 1M tokens

2 routes · 1 cache

Xiaomi 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.

$0.140 in / $0.280 out / 1M tokens

2 routes · 2 cache

Xiaomi 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.

$0.140 in / $0.280 out / 1M tokens

2 routes · 1 batch · 2 cache

Xiaomi 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.

$0.435 in / $0.870 out / 1M tokens

2 routes · 1 batch · 2 cache

Kimi 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.

$3.00 in / $15.00 out / 1M tokens

3 routes · 2 cache

Hermes 4's 405B hosted variant on Nous Portal. The portal describes it as the largest Hermes 4 model, focused on advanced reasoning and creative depth rather than inference speed or cost.

$0.090 in / $0.370 out / 1M tokens

2 routes

Nanbeige4-3B is an open-source 3B-parameter language model by Nanbeige LLM Lab (BOSS Zhipin), released December 2025. Pre-trained on 23 trillion high-quality tokens with SFT on 30M+ diverse instructions. Context extended to 64K via Adjusted Base Frequency (ABF). Sets state-of-the-art on AIME 2024 (90.4), AIME 2025 (85.6), and GPQA-Diamond (82.2) for sub-10B models, outperforming models up to 10× larger including Qwen3-32B. arxiv: 2512.06266. HuggingFace: Nanbeige/Nanbeige4-3B-Base.

No tracked provider route

GLM-5.3-Flash is Z.ai's first native multimodal model in the GLM-5 series. It is a 320B-total / 18B-active Mixture-of-Experts model with a 1M-token context window and 131,072 max output tokens, supporting image, video, and file input plus function calling, tool use, structured outputs, reasoning, and prompt caching. MIT open weights are ungated on Hugging Face at zai-org/GLM-5.3-Flash. Official API model ID: glm-5.3-flash. The same model was previously tested anonymously as ox-alpha; that name is not a separate catalog slug.

$0.150 in / $0.500 out / 1M tokens

2 routes · 2 cache

GLM-5.3 is Z.ai's August 2026 coding-first post-training release on the same GLM-5.2 base model, focused on complex software engineering, long-horizon agentic coding, and emergent cybersecurity evaluation. It is available through the GLM Coding Plan and Z.ai API; open weights are scheduled after safety hardening. Official API model ID: glm-5.3.

$1.40 in / $4.40 out / 1M tokens

3 routes · 2 cache

GLM-5.2 is Z.ai's coding-first successor to GLM-5.1 in the GLM-5 family, released June 13 2026. 753B parameters (40B active) in IndexShare MoE architecture; the IndexShare innovation reuses the same attention indexer across every four sparse layers, cutting per-token FLOPs by 2.9x at 1M context length. Trained on 28.5T tokens. Supports a 1M-token context window via the glm-5.2[1m] model ID, with 131,072-token maximum output and High/Max thinking-effort levels designed for extended agentic coding sessions. MIT license; open weights available on Hugging Face (zai-org/GLM-5.2 and zai-org/GLM-5.2-FP8). Self-reported HF card benchmarks: SWE-bench Pro 62.1, Terminal-Bench 2.1 82.7, MCP-Atlas 76.8, Tool-Decathlon 48.2, GPQA Diamond 91.2, AIME 2026 99.2, HLE 40.5. Available to GLM Coding Plan subscribers (Lite/Pro/Max/Team) directly, and via OpenRouter token API ($1.40/$4.40 per 1M tokens).

$1.40 in / $4.40 out / 1M tokens

2 routes · 1 cache

Grok-0, developed by xAI, is a large language model (LLM) that boasts 33 billion parameters 237. It impressively performs on par with Meta's 70-billion parameter LLaMA 2 model, despite utilizing only half the training resources, highlighting its architectural efficiency and optimization 235. Grok-0 served as the prototype for this efficient design before being succeeded by Grok-1, which further enhanced reasoning and coding capabilities 2.

No tracked provider route

Grok-1, created by xAI, is a formidable 314-billion parameter Mixture-of-Experts (MoE) language model. It boasts a sophisticated architecture with 8 experts, leveraging 2 for each token input, spread across 64 layers and equipped with 48 attention heads per query. This vast model was trained from scratch using a specially crafted training stack based on JAX and Rust, finishing its pre-training phase by October 2023. Released as a base model under the permissive Apache 2.0 license, its open-source framework allows both commercial and non-commercial applications, though it lacks fine-tuning for specific tasks. Benchmarks highlight Grok-1's superior reasoning on various tasks but recognize its potential for generating inaccuracies ("hallucinations"). Running on a local setup requires substantial hardware, including a multi-GPU system, for efficient performance.

No tracked provider route

Grok-1.5V, created by xAI, is a multimodal large language model that combines both text and image processing capabilities. This model excels at interpreting and interacting with diverse visual data, including documents, diagrams, charts, screenshots, and photographs. Its multimodal nature allows it to perform advanced tasks like translating diagrams into code, generating image descriptions, and answering questions based on visual inputs, all while displaying a strong understanding of spatial information. Grok-1.5V has demonstrated competitive prowess against top models such as GPT-4V and Gemini Pro 1.5, particularly in areas that require spatial reasoning. Initially, access is primarily limited to early testers and existing Grok users, with plans for broader availability in the future 124.

No tracked provider route

Grok-1.5, developed by xAI, Elon Musk's AI company, is a large language model focused on advanced reasoning skills in coding and mathematics, highlighted by its exceptional performance on benchmarks such as MATH, GSM8K, and HumanEval. It supports handling long contexts of up to 128,000 tokens, surpassing its predecessor in this area, and is built using a custom distributed training framework on JAX, Rust, and Kubernetes. Designed for comprehensive context understanding and logical reasoning, it is being deployed to early testers and users. Additionally, a multimodal version, Grok-1.5V, is available, which incorporates visual information processing capabilities, including documents, diagrams, and photographs.

No tracked provider route

Post-training variant of GLM-5 from Z.ai (Zhipu AI) with enhanced agentic coding capabilities. Released April 7, 2026. 754B parameters (40B active) in Mixture of Experts architecture, 200K token context, 128K max output. Supports autonomous plan–execute–test–fix–optimize loops for up to 8 hours without human intervention. Trained entirely on Huawei Ascend hardware (no Nvidia). Key benchmarks: SWE-bench Pro 58.4 (world #1 at release, surpassing GPT-5.4 57.7 and Claude Opus 4.6 57.3), GPQA Diamond 86.2, AIME 2026 95.3, Terminal-Bench 2.0 63.5, MCP-Atlas 71.8, Chatbot Arena Elo 1475 (June 16, 2026, arena.ai). Available via Z.ai API ($1.40/$4.40 per 1M input/output tokens) and open weights on Hugging Face under MIT license.

$1.38 in / $4.40 out / 1M tokens

5 routes · 3 cache

Lightweight variant of Grok-2 from xAI with extended context for general reasoning tasks.

No tracked provider route

Grok-2 is xAI's Grok 2 model with an optional reasoning mode. It offers a 128K-token context window.

$0.500 in / $0.500 out / 1M tokens

1 route

Amazon Nova Premier is Amazon's most capable standard Bedrock Nova understanding model for complex reasoning, agentic workflows, and model distillation. It supports a 1M-token context window, text/image/video inputs, text output, reasoning, tool calling, and prompt caching; use it as the standard Bedrock Nova frontier pick instead of Nova 2 Omni early-access Forge checkpoints.

$2.50 in / $12.50 out / 1M tokens

2 routes · 1 batch · 2 cache

DeepSeek R1: Reasoning-optimized model with extended thinking capabilities. 128K context.

$0.100 in / $0.300 out / 1M tokens

9 routes

NVIDIA's open frontier-reasoning model (550B total / 55B active MoE, hybrid Transformer-Mamba). 300+ tokens/second. 1M-token context. Announced at Computex 2026. Pricing: ~$0.60/$2.60 per 1M tokens (provider median); free tier on some providers.

$0.900 in / $1.70 out / 1M tokens

8 routes · 6 cache

Flagship open-weight foundation model from Zhipu AI with 744B parameters (40B active per token) in Mixture of Experts architecture. Trained on 28.5T tokens using DeepSeek Sparse Attention on Huawei Ascend hardware. Achieves state-of-the-art performance on coding and agentic benchmarks (SWE-bench Verified: 77.8%). Supports autonomous planning, multi-step tool use, and self-correction.

$1.00 in / $3.20 out / 1M tokens

7 routes · 3 cache

Kimi 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.

$2.00 in / $10.00 out / 1M tokens

6 routes · 5 cache

DeepSeek V4 Pro is DeepSeek's flagship open-weights model, released April 24 2026 under the MIT license. Architecture: 1.6T total / 49B active parameters, MoE with Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) hybrid — requiring only 27% of inference FLOPs vs standard 1M-context transformers — plus Manifold-Constrained Hyper-Connections (mHC) and Muon Optimizer. Context window: 1,000,000 tokens; max output: 384,000 tokens (Think Max mode requires >=384K context). Text-only (no vision/image input). Supports three reasoning modes: Non-Think, Think High, Think Max. Function calling, tool use, and structured outputs supported. Key benchmarks: SWE-bench Verified 80.6%, SWE-bench Pro 55.4%, LiveCodeBench 93.5%, GPQA Diamond 90.1%, MMLU-Pro 87.5% and Chatbot Arena 1456 (2026-06-16).

$0.3633 in / $0.7266 out / 1M tokens

6 routes · 4 cache

NVIDIA 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.

$0.085 in / $0.400 out / 1M tokens

6 routes · 2 cache

Kimi K2 Instruct is an instruction-tuned language model from Moonshot AI, available via Fireworks AI.

$0.570 in / $2.30 out / 1M tokens

6 routes

Alibaba's closed-weight flagship language model, announced at the 2026 Alibaba Cloud Summit (May 20). 1M-token context with prompt caching.

$1.25 in / $3.75 out / 1M tokens

4 routes · 2 cache

Muse 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.

$0.100 in / $0.500 out / 1M tokens

3 routes · 3 cache

Particle-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.

$0.250 in / $1.00 out / 1M tokens

3 routes · 3 cache

GLM-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.

$0.370 in / $1.25 out / 1M tokens

3 routes · 3 cache

DeepSeek R1 Distill Llama 70B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.800 in / $0.800 out / 1M tokens

3 routes

DeepSeek R1 Distill Qwen-14B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.150 in / $0.150 out / 1M tokens

3 routes

DeepSeek R1 Distill Qwen-32B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.900 in / $0.900 out / 1M tokens

3 routes

LongCat-2.5 is Meituan's native multimodal model for long-horizon agentic tasks, released November 2, 2026. It is a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token, built on LongCat-2.0 with stronger native visual understanding and agentic coding, with W4A4 inference optimized for domestic AI accelerators. It accepts text and image input. It is served on the LongCat API Platform as LongCat-2.5, and Meituan released the weights on Hugging Face under the MIT License on November 9, 2026.

Pricing not tracked / 1M tokens

2 routes

Kimi 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.

$4.00 in / $20.00 out / 1M tokens

2 routes · 2 cache

Ampersand (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.

$1.00 in / $5.20 out / 1M tokens

2 routes · 1 cache

Qwen3.8-Max is a 2.4-trillion-parameter MoE flagship delivering a comprehensive leap in coding and professional work. It autonomously codes and delivers complete projects spanning 10+ days, handles hundreds of specialized tasks across legal, financial, design, and other professional domains, and produces production-grade results end-to-end in a single conversation. Native visual understanding runs through the full cycle of planning, execution, and verification, enabling deep semantic analysis of ultra-long documents and extended video content. The public Hugging Face checkpoint (Qwen3.8-2.4T-A95B) is text-only under the Qwen3.8-Max License; the hosted qwen3.8-max API adds vision, optional non-thinking mode, 1M context, tools, structured outputs, and prompt caching on the same model ID.

$2.00 in / $6.00 out / 1M tokens

2 routes · 2 cache

DeepSeek R1 Distill Llama 8B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.200 in / $0.200 out / 1M tokens

2 routes

DeepSeek R1 Distill Qwen-7B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.200 in / $0.200 out / 1M tokens

2 routes

K-EXAONE 2.5 is LG AI Research's proprietary Korean-English language model, released November 9, 2026 as a technical validation model further trained from K-EXAONE 2.0. It has 750B total parameters (31B active), a 524,288-token context and a January 2026 knowledge cutoff, and supports chat, tool calling, structured outputs and code generation. It is served on the EXAONE API Platform as k-exaone_v2.5.

Pricing not tracked / 1M tokens

1 route

SWE-2 Flash is Cognition's fast, low-cost coding model for Devin, released November 4, 2026. It is post-trained from Thinking Machines' Inkling-Small (276B total, 12B active) with a mid-training stage added before reinforcement learning, is available in Devin Desktop and CLI. Cognition is building a program to share the weights securely with industry partners and researchers.

Pricing not tracked / 1M tokens

1 route

Ampersand Tutor (&Tutor) is humans&'s tutoring model, released October 30, 2026 through the Ampersand API as ampersand-tutor-v0.1. Built on GLM 5.3, it learns to teach by training against Persimmon, humans&'s large-scale model of how students learn, and aims to maximize a student's understanding rather than just give answers. It is a mixture-of-experts transformer (78 layers, 256 routed experts plus 1 shared, 8 active per token) served in NVFP4, with text input and output.

$1.00 in / $5.20 out / 1M tokens

1 route · 1 cache

Riverflow 2.6 Pro is Sourceful's flagship image generation and editing model, released October 29, 2026 as the upgrade to Riverflow 2.5 Pro. It generates and edits product and campaign images from text and reference images, supports 1K, 2K and 4K output with more than 46 million size combinations, and has visual reasoning built in. It is available in the Riverflow app, MCP and API.

Pricing not tracked / 1M tokens

1 route

Flint is Recursive Superintelligence's first model, released October 21, 2026 as open weights under Apache 2.0. It is a mixture-of-experts coding model with 39.8B total and about 2.5B active parameters per token, trained from scratch and built for agentic coding on local hardware: investigating codebases, fixing bugs, implementing features and writing tests, with always-on reasoning and multi-turn tool use. Its documented context configuration is 262,144 tokens (the blog says up to 512k is supported) and Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. Recursive says it runs at more than 125 tokens/s on an M4 Pro MacBook Pro and more than 150 tokens/s on a 64GB NVIDIA RTX Spark at short contexts, with quantized INT4 and NVFP4 checkpoints for Apple Silicon and NVIDIA runtimes.

Free in / Free out / 1M tokens

1 route

Cosmos 3 Nano is NVIDIA's 16B-parameter omnimodel optimized for efficient inference on workstation-grade hardware (NVIDIA RTX PRO 6000). Architecture: dual-tower Mixture-of-Transformers with an 8B autoregressive Reasoner and an 8B diffusion-based Generator. The Reasoner supports up to 256K tokens of context for vision-language reasoning; the Generator produces video up to 720p at variable frame rates (default 189 frames). Natively handles text, image, video, audio (48kHz stereo), and robot action trajectories across 10+ robot embodiments including Franka Panda, UR, Google robot, and UMI. BF16 precision only. Available as open weights on Hugging Face and via the Cosmos 3 Reasoner NIM (NIM_MODEL_SIZE=nano). Intended for real-time robotics inference and edge-adjacent deployment. Robot action input/output is preserved in this description because the model schema does not have a dedicated action modality field.

Pricing not tracked / 1M tokens

1 route

Cosmos 3 Super is NVIDIA's flagship 64B-parameter omnimodel for physical AI, designed for large-scale synthetic data generation and high-fidelity simulation on NVIDIA Hopper and Blackwell datacenter GPUs. Architecture: dual-tower Mixture-of-Transformers with a 32B autoregressive Reasoner and a 32B diffusion-based Generator. Supports 256K token reasoning context, 720p video generation at variable frame rates, and 10+ robot embodiment action domains. Ranked #1 among open models on Physics-IQ, PAI-Bench, R-Bench, RoboLab, RoboArena, VANTAGE-Bench, and TAR (Computex 2026). Training data: 1.3B data points across 393 datasets (2024-2026). Inference performance (vLLM-Omni): ~55s for 50-step video on 8xH200. Available as open weights on Hugging Face and via Cosmos 3 Reasoner NIM (NIM_MODEL_SIZE=super).

Pricing not tracked / 1M tokens

1 route

Perceptron Mk1 is a closed-source vision-language model for image and video understanding, OCR, object detection, captioning, video QA, and embodied reasoning. Perceptron documents Mk1 with 32K context, reasoning support, and standard pricing of $0.15 per 1M input tokens and $1.50 per 1M output tokens.

$0.150 in / $1.50 out / 1M tokens

1 route

Hermes 4's 70B hosted variant on Nous Portal. The portal describes it as a hybrid-mode reasoning model that balances scale and size while staying fast and cost effective for complex reasoning tasks.

$0.050 in / $0.200 out / 1M tokens

1 route

MiniMax-M1 is a large-scale open-weight reasoning model from MiniMax with 456B total parameters and a 1M token context window, designed for extended reasoning and high-efficiency inference.

$0.400 in / $2.20 out / 1M tokens

1 route

DeepSeek R1 Distill Qwen-1.5B is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

$0.100 in / $0.100 out / 1M tokens

1 route

Spark 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.

No tracked provider route

Nanbeige4.3-33B is the Nanbeige team's 33B dense open-weights language model, a post-trained model on Nanbeige4.3-33B-Base with a 512K-token (524,288) context window, and a switchable thinking mode (enable_thinking). Released 2026-10-25 under Apache 2.0.

No tracked provider route

Ling-3.1-flash-1019 is inclusionAI's (Ant Group Ling team) upgraded checkpoint of Ling-3.1-flash: a hybrid-linear mixture-of-experts model with 560B total and 25B active parameters per token (512 routed experts, 8 active). Further post-training improves reasoning and real-world agentic work. Announced 2026-10-22 and open-sourced on Hugging Face on 2026-10-27 under the MIT license.

No tracked provider route

Sage-A1.5 is Shanghai AI Laboratory's 35B MoE agentic model built on Qwen3.6-35B-A3B and Agents-A1 for long-horizon scientific problem solving, adding agentic orchestration of subagents, verification and open scientific discovery. 256K native context (512K with YaRN); Apache 2.0.

No tracked provider route

Spark 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.

No tracked provider route

PhoneLLM Max Alpha 2 is the larger variant of Daily's Pipecat PhoneLLM, an open-weights text model for voice agents. A full-parameter supervised fine-tune of the language backbone of NVIDIA Nemotron 3.5 Super VL (120B total, 12B active parameters), it emits a single sentence of reasoning ('microthinking') before each reply to improve tool-call accuracy, and supports Pipecat Flows. Released 2026-10-15 under the OpenMDW 1.1 license.

No tracked provider route

MiniCPM-V 4.7 1B is OpenBMB's lightweight on-device vision-language model for single-image, multi-image and video understanding. It compresses visual tokens early inside the vision encoder (LLaVA-UHD v4 technique) with mixed 4x/16x compression, runs on iOS, Android and HarmonyOS, and supports Instruct and Thinking modes. OpenBMB reports it improves on MiniCPM-V 4.6 and beats Qwen3.5 0.8B. Apache 2.0.

No tracked provider route

MiniCPM-V 4.7 35B-A3B is the first mixture-of-experts model in OpenBMB's MiniCPM-V series (~3B active parameters), built for on-device multimodal use: with Lookahead Expert Scheduling it decodes at 13+ tokens/s in about 3 GB of memory on iPhone 17 Pro Max, per OpenBMB. It handles images, multi-image and video, with Instruct and Thinking modes and tool calling. Apache 2.0.

No tracked provider route

1-bit Bonsai 27B is PrismML's footprint-oriented 27B model from the July 2026 Bonsai 27B launch: Qwen3.6 27B compressed to binary weights at 1.125 effective bits per weight, about 3.9 GB, small enough for an iPhone 17 Pro. PrismML reports it keeps about 90% of full-precision performance across 15 thinking-mode benchmarks, with reasoning, tool calling and optional image input; 262K-token context. Sibling of Ternary Bonsai 27B. Apache 2.0.

No tracked provider route

Ring-2.6-1T is InclusionAI's MIT-licensed trillion-parameter MoE reasoning model for agent workflows, engineering tasks, scientific analysis, and enterprise automation. It supports high and xhigh reasoning effort modes and entered OpenRouter's Programming top 10 in the 2026-05-18 audit.

No tracked provider route

CoBuddy is a Baidu Qianfan code generation model optimized for coding tasks and AI agent workflows. OpenRouter lists the free variant with a 131K context window, native tool support, reasoning support, and FP8 quantization for high-throughput inference.

No tracked provider route

Nanbeige4.1-3B is an open-source 3.93B-parameter reasoning and agentic language model by Nanbeige LLM Lab (BOSS Zhipin), released February 11, 2026. Built on Nanbeige4-3B-Base with further SFT and reinforcement learning. Supports 256K token context. A unified generalist model achieving strong reasoning, preference alignment, and agentic tool-use capabilities at the ~3B scale. Competitive with much larger open models. arxiv: 2602.13367. HuggingFace: Nanbeige/Nanbeige4.1-3B.

No tracked provider route

Tencent Hunyuan T1 is a deep-thinking reasoning model released March 21, 2025. Built on Hunyuan TurboS base, using a Hybrid-Transformer-Mamba MoE architecture — the first ultra-large-scale Mamba-powered LLM with 16 total experts and 52B activated parameters via dynamic routing. 96.7% of compute allocated to reinforcement learning post-training. Benchmarks: MATH-500 96.2, LiveCodeBench 64.9, GPQA Diamond 69.3, MMLU-PRO 87.2 (second only to o1). Decoding speed 2× faster than comparable transformer models. 256K token context. Available on Tencent Cloud API. Source: https://tencent.github.io/llm.hunyuan.T1/README_EN.html

No tracked provider route

DeepSeek R1 Zero is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

No tracked provider route

DeepSeek R1 Lite is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 128K-token context window with weights openly available for self-hosting.

No tracked provider route

The Cerebras GPT 590M is a robust language model featuring 590 million parameters and a transformer architecture akin to GPT-3. It is optimized for natural language processing tasks such as text generation, completion, and summarization. Trained using the Chinchilla scaling laws and Cerebras' weight streaming technology, this model achieves high efficiency, offering faster training times and reduced costs. The Andromeda AI supercomputer facilitated its training on the extensive Pile dataset. Open-sourced under the Apache 2.0 license, it primarily supports English and requires additional tuning for other languages and conversational applications due to its lack of reinforcement learning from human feedback.

No tracked provider route

The NeMo Megatron-GPT 5B is a transformer-based language model with 5 billion trainable parameters, inspired by models like GPT-2 and GPT-3 1. Its architecture is a decoder-only transformer, designed to sequentially process input for text generation and language understanding tasks 15. Trained on "The Piles" dataset by Eleuther.AI, it leverages its substantial dataset to produce coherent and natural-sounding text while also answering questions and completing sentences 5. Despite its strengths, the model can reflect biases and toxic language from its dataset, sometimes yielding inappropriate outputs. Evaluations on benchmarks like the LM Evaluation Test Suite showcase its varying performance, scoring 0.5566 on ARC-Easy and 0.6133 on Winogrande 1, indicating both strengths and limitations across different tasks.

No tracked provider route

OpenAI's previous intelligent reasoning model with configurable reasoning effort. Released August 2025. Supports minimal, low, medium, and high reasoning levels. Succeeded by GPT-5.1 and later models.

$1.25 in / $10.00 out / 1M tokens

5 routes · 1 batch · 3 cache

Near-frontier intelligence for cost-sensitive, low-latency, high-volume workloads. Released August 2025. Replaces o4-mini (shutting down Oct 2026).

$0.250 in / $2.00 out / 1M tokens

4 routes · 1 batch · 3 cache

Fastest, cheapest GPT-5 variant for summarization and classification tasks. Also available via Realtime API.

$0.050 in / $0.400 out / 1M tokens

4 routes · 1 batch · 3 cache

Premium 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.

$30.00 in / $180.00 out / 1M tokens

3 routes · 1 batch

Advanced o3 reasoning model for complex math, science, and coding problems. Supports tools, vision, and extended thinking. Available to Pro users. Released June 10, 2025.

$20.00 in / $80.00 out / 1M tokens

3 routes

Seed 1.6 Flash is ByteDance Seed's ultra-fast multimodal thinking model supporting text and visual understanding at 256K context, optimized for low-latency inference.

$0.075 in / $0.300 out / 1M tokens

1 route

LFM2.5-24B-A2B is Liquid AI's largest LFM2.5 model: a 24B-parameter mixture-of-experts reasoning model with about 2B active parameters per token, built for fast on-device agents. It keeps the LFM2 hybrid of gated short convolutions and GQA (40 layers, 64 experts, top-4 routing), extends context from 32,768 to 131,072 tokens, doubles the vocabulary to 128K, and scales pre-training from 17T to 33T tokens. Unlike LFM2-24B-A2B, it is reasoning-only (explicit chain of thought before the final answer) and is post-trained for tool calling and multi-step agent harnesses. Open weights under LFM Open License v1.0 on Hugging Face; also on Liquid Console, OpenRouter, and Vercel AI Gateway.

$0.050 in / $0.700 out / 1M tokens

3 routes · 3 cache

Seed 1.6 is a general-purpose multimodal model from ByteDance Seed supporting text, image, and video inputs. It incorporates multimodal capabilities and deep thinking for complex tasks at 256K context.

$0.250 in / $2.00 out / 1M tokens

2 routes · 1 cache

Amazon Nova 2 Lite is a fast, cost-effective reasoning model for everyday workloads that processes text, images, and videos at 1M token context with improved reasoning over Nova Lite v1.

$0.300 in / $2.50 out / 1M tokens

2 routes · 1 cache

MAI-Thinking-1 is Microsoft AI's flagship reasoning model, built from scratch on enterprise-grade commercially licensed data without third-party distillation. The sparse mixture-of-experts model activates about 35B parameters from roughly 1T total parameters, supports a 256K-token context window, and targets frontier reasoning and software engineering work at a mid-weight price point. Microsoft reports 97% on AIME 2025, 94.5% on AIME 2026, 84.2% on GPQA Diamond, 87.7% on LiveCodeBench v6, 73.5% on SWE-bench Verified, and 52.8% on SWE-bench Pro. In a 1,276-task Surge blind side-by-side evaluation, it narrowly beat Claude Sonnet 4.6 but trailed Claude Opus 4.6. It supports function calling and developer instructions through the Chat Completions API.

$2.00 in / $8.00 out / 1M tokens

1 route · 1 cache

Qwen3.8-Flash-Next is Alibaba's experimental open-weight preview of the architecture planned for Qwen4. It is a 125B-total / 6B-active Mixture-of-Experts causal language model with a vision encoder, plus 51B n-gram embedding and 4B MTP parameters. Native context is 262,144 tokens (extensible to 1,000,000). Weights are on Hugging Face under the Qwen Community License 1.0.

No tracked provider route

Open-weight dense Qwen3.6 27B model with native multimodal support across text, image, and video. Apache 2.0.

$0.320 in / $2.70 out / 1M tokens

5 routes