Best LLMs for Agents (2026)

Anthropic's most capable widely released model, built for demanding reasoning and long-horizon agentic work. Claude Fable 5 is the generally available Mythos-class Claude model, supports vision, tool use, structured outputs, prompt caching, Batch API processing, adaptive thinking that is always on, a 1M-token context window, and up to 128k output tokens. It launched on the Claude API, AWS Bedrock, Vertex AI, and Microsoft Foundry on June 9, 2026 with first-party pricing at $10 per 1M input tokens and $50 per 1M output tokens. Anthropic temporarily disabled Fable 5 access on June 12, 2026 after a US export control directive and redeployed Fable 5 and Mythos 5 on July 1, 2026.

Claude Opus 5 is Anthropic's July 2026 flagship Opus model for complex agentic coding, enterprise work, long-horizon reasoning, computer use, and professional analysis. It accepts text and image input and returns text, with adaptive thinking enabled by default and five request-level effort settings from low through max that control thinking depth and token use rather than visible response length. It also supports tool use, structured outputs, prompt caching, Batch API, a 1M-token context window, and up to 128K synchronous output tokens. Anthropic lists standard Claude API pricing at $5 per million input tokens and $25 per million output tokens, with Fast mode at $10/$50. It is available through the Claude API and hosted provider routes.

Claude Opus 4.8 is Anthropic's flagship Claude 4.8 model, released May 28, 2026 for agentic coding, long-horizon reasoning, computer use, and professional knowledge work. It supports text and image inputs, adaptive reasoning, tool use, structured outputs, computer-use tools, prompt caching, Batch API, Dynamic Workflows parallel subagents, a 1M-token context window on Anthropic API/Bedrock/Vertex, and 128K max output. Reported benchmark scores: SWE-bench Pro 69.2%, SWE-bench Verified 88.6%, Terminal-Bench 2.1 74.6%, HLE with tools 57.9%, OSWorld-Verified 83.4%, GDPval-AA 1890 Elo, and MCP-Atlas 82.2%. Standard Anthropic API pricing is $5/M input and $25/M output.

Claude Opus 4.7 is Anthropic's generally available flagship model with 1M context, 128K max output, adaptive thinking, and a new tokenizer with roughly 555K words per 1M tokens.

Claude Sonnet 5 is Anthropic's next-generation Sonnet model for agentic coding, tool use, computer use, and professional work. It is a proprietary decoder-only model with a 1M-token context window, 128K max output, multimodal vision, adaptive thinking, function calling, structured outputs, prompt caching, and Batch API support. It is available through the Claude API, AWS Bedrock, Google Cloud Vertex AI, Microsoft Foundry preview, and OpenRouter. Anthropic lists durable standard pricing at $3/1M input and $15/1M output tokens, with introductory $2/$10 pricing through 2026-08-31.

Most capable agentic coding model from OpenAI. Optimized for long-horizon, agentic coding tasks in the Codex CLI and API. Note: GPT-5.3-Codex-Spark is a distinct ChatGPT Pro research preview (not API-accessible).

GPT-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.

GPT-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.

Beam 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).

Claude Opus 4.5 is Anthropic's Claude 4.5 model with multimodal text and image input and an optional reasoning mode. It offers a 200K-token context window and scores 80.7 on MMMU.

Claude Opus 4.6 is Anthropic's Claude 4.6 model with multimodal text and image input and an optional reasoning mode. It offers a 1M-token context window and scores 80.8 on SWE-bench Verified.

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%, Terminal-Bench 2.0 59.1% on BenchLM's independent June 2026 harness, and Chatbot Arena 1456 (2026-06-16). Current API pricing: $0.435/$0.87 per 1M input/output tokens; DeepSeek made the former 75% promotional rate permanent in May 2026.

Google: Gemini 3.1 Pro Preview available via OpenRouter. Pricing: $2/1M input, $12/1M output.

MiniMax M3 is MiniMax's current API flagship (released June 1, 2026) with MiniMax Sparse Attention for economical 1M-token context, native multimodality, and agentic coding. Benchmark rows in LLMReference separate vendor-reported SWE-bench Pro, Terminal-Bench 2.1, MCP-Atlas, and BrowseComp scores (source: minimax.io) from third-party rows — inspect evaluator, variant, and source on each benchmark cell before comparing to leaderboard claims.

Alibaba's closed-weight flagship language model, announced at the 2026 Alibaba Cloud Summit (May 20). Scored 56.6 on Artificial Analysis Intelligence Index at launch—highest-ranked Chinese model. 1M-token context with prompt caching (up to 90% discount). Pricing: $2.50/$7.50 per 1M tokens in/out.

Kimi K2.6 is Moonshot AI's multimodal agentic coding model, released April 20 2026 under a Modified MIT license. Built on a 1-trillion-parameter MoE architecture (32B active, 384 experts with 8 selected per token plus 1 shared expert, 61 layers), it features a 262K context window and up to 65,536 output tokens. Supports native image and video inputs (screenshots, PDFs, spreadsheets). Designed for long-horizon coding with agent swarms of up to 300 sub-agents and 4,000 coordinated steps; Moonshot AI cites 200–300 sequential tool calls without task drift. Key benchmarks: SWE-bench Verified 80.2%, SWE-bench Pro 58.6%, LiveCodeBench v6 89.6%, GPQA Diamond 90.5%, Terminal-Bench 2.0 66.7%. Chatbot Arena Elo 1454 (2026-04-28 snapshot).

MiniMax M2.5 Highspeed is MiniMax's inference-optimized variant of M2.5, released simultaneously in February 2026. It delivers identical intelligence and outputs to standard M2.5 through a specialized inference engine at lower latency. The model supports a 204,800-token context window, 131,072-token max output, function calling, structured output, and reasoning. API model ID: MiniMax-M2.5-highspeed. It is designed for latency-sensitive interactive applications and automated agent pipelines.

Claude Sonnet 4.6 is Anthropic's best combination of speed and intelligence. Proprietary decoder-only model with 1M-token context, 64K max output, multimodal vision, extended thinking, and function calling. Available via Anthropic API, AWS Bedrock, GCP Vertex AI, and OpenRouter at $3/1M input and $15/1M output tokens.

DeepSeek V4 Flash is a 284B parameter (13B activated) Mixture-of-Experts language model with 1M-token context. It uses a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for long-context inference, and supports thinking and non-thinking modes. MIT licensed. The legacy API aliases deepseek-chat and deepseek-reasoner mapped to its non-thinking and thinking modes and were scheduled for discontinuation on July 24, 2026. DeepSeek retired the model from its API on September 10, 2026. The legacy model name deepseek-v4-flash is still accepted, and DeepSeek now serves those requests with DeepSeek-V4.1-Flash-1017.

Xiaomi'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.

Qwen3.6-Plus is Alibaba Cloud's GA Qwen3.6 flagship for long-context reasoning, coding, tool use, and multimodal workflows. DashScope lists it with a 1M-token context window, structured output support, and standard public token pricing.

Gemini 3 Flash is Google's speed-optimized Gemini 3 model, available in public preview via the Gemini API and Vertex AI. It supports text, image, audio, and video inputs with a 1M token context window and is priced at $0.50 per 1M input tokens and $3.00 per 1M output tokens.

Gemini 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.

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.

Mistral's 128B flagship merged model, released April 29, 2026. Open weights available on HuggingFace since ~May 22, 2026 under a Modified MIT License (free for companies with <$20M/month revenue). Dense 128B architecture with 256k context window, configurable reasoning effort, vision encoder for variable image sizes, and native function calling. Achieves 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified. Strong multilingual support across 24+ languages. Replaces Mistral Medium 3.1, Magistral, and Devstral 2 as Mistral's primary merged model. Available via Mistral API at $1.50/$7.50 per million tokens (input/output) and via NVIDIA NIM/endpoints.

Muse Spark is the first model in Meta's Muse family, developed by Meta Superintelligence Labs (MSL). It is a natively multimodal reasoning model with capabilities including tool-use, visual chain-of-thought reasoning, and multi-agent orchestration. Muse Spark achieves 58% on Humanity's Last Exam and 38% on FrontierScience Research benchmarks, while being competitive with Llama 4 Maverick at over 10x less compute. Available via meta.ai and the Meta AI app; private API preview only — not open-source.

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

Claude Sonnet 4.5 is Anthropic's Claude 4.5 model with multimodal text and image input and an optional reasoning mode. It offers a 200K-token context window and scores 86 on MMLU PRO.

DeepSeek: DeepSeek V3 0324 available via OpenRouter. Pricing: $0.2/1M input, $0.77/1M output.

Kimi K2.5 is Moonshot AI's Kimi model focused on code generation and software engineering. It offers a 256K-token context window and scores 87.9 on GPQA.

Original Grok 4 API model (grok-4-0709) with 256K context, reasoning, text and image input, function calling, structured outputs, live browsing, and code execution support. xAI lists May 15, 2026 as the API retirement date for this model; new integrations should use Grok 4.3.

Grok 4.20 is xAI's February 2026 Grok 4-series model, first previewed under the informal Grok 4.2 beta label. Standard API variants launched around March 10, 2026 as grok-4.20-0309-reasoning and grok-4.20-0309-non-reasoning with a 1M context window.

Doubao Seed 2.0 Pro is ByteDance's flagship frontier model released February 14, 2026 via Volcano Engine. Achieves 88.9 on GPQA Diamond, 87.0 on MMLU-Pro, 87.8 on LiveCodeBench v6, and 76.5 on SWE-bench Verified. Priced at $0.47/M input and $2.37/M output, making it 3–6× cheaper than comparable frontier models.

Doubao Seed 2.0 Code is ByteDance's coding-specialized frontier model released February 14, 2026 via Volcano Engine. Achieves 87.8 on LiveCodeBench v6 and 76.5 on SWE-bench Verified. Tuned for software development tasks including code generation, debugging, repository-level understanding, and automated code review.

Qwen3.5-Plus is the flagship commercial API model of the Qwen3.5 native vision-language series, delivering outstanding performance comparable to state-of-the-art models with significant leaps in both pure-text and multimodal capabilities compared to the Qwen3 series.

Alibaba's largest Qwen3.5 model, featuring a Mixture-of-Experts architecture with 397B total parameters and 17B active per token (using 512 total experts with 10 routed + 1 shared active). Supports 201 languages with a native 262K token context window extensible to 1M tokens via YaRN. Includes a thinking/reasoning mode, tool calling with MCP integration, and unified vision-language capabilities through early fusion training.

Motif 3 (HF Motif-Technologies/Motif-3) is Motif Technologies' decoder-only mixture-of-experts LLM with ~314B total and 13.2B active parameters, trained from scratch on ~12.5T tokens with extra Korean, reasoning and agentic data. It uses Grouped Differential Latent Attention (GDLA), Expert-Specific PolyNorm, modified mHC and a built-in MTP head for self-speculative decoding. Native context is 262,144 tokens. It targets long-horizon agentic tool use and terminal tasks, with reasoning and tool-call parsers for vLLM. MIT License; base checkpoint Motif-Technologies/Motif-3-Base.

Muse Glimmer is Meta Superintelligence Labs' 30-billion-parameter open-weight multimodal agentic model for local deployment, released August 10, 2026 under Apache 2.0. Distilled from Muse Spark, it combines a ~2B perception encoder with a ~28B text decoder for tool use, multimodal reasoning, and long-horizon agent workflows on consumer GPUs. Weights are available on Hugging Face at meta-models/Muse-Glimmer-30B. Meta's technical report (October 19, 2026) describes it as a dense causal transformer with about 29.6B total parameters (about 1.8B in the frozen perception encoder), a 128K+ context window and four reasoning strengths (low, medium, high, xhigh). Meta publishes two roughly 4-bit builds, a 16.8 GB K-Quant for 24 GB GPUs and a 19.7 GB K-Quant Dynamic for 32 GB GPUs, plus a DFlash speculative-decoding drafter; the report says it runs at practical speeds on a single RTX 5090. The BF16 weights take 55.7 GB.

Frontier-class performance rivaling larger models at a fraction of the cost. Most intelligent Gemini model built for speed, combining frontier intelligence with superior search and grounding. $0.50 input / $3.00 output per 1M tokens.

Unisound U2 is a native agentic large language model announced June 8, 2026, by Chinese AI company Unisound (云知声). Designed for multi-step autonomous task execution, it uses native reasoning-path distillation for task decomposition, execution, verification, and optimization. The official Token Hub page lists a sparse MoE architecture with 266B total parameters and 10B active parameters, plus API access through Unisound Token Hub. Emphasis is on coding, tool orchestration, and long-horizon agent workflows. Unisound self-reports GPQA Diamond 87.9, SWE-bench Verified 75.0, and Claw-Eval 76.9; treat those benchmarks as vendor-reported until independently verified. Context window is not disclosed.

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.

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.

Step 3.5 Flash is StepFun's most capable open-source foundation model. Its sparse Mixture of Experts architecture activates only 11B of 196B total parameters per token, making it highly efficient at long contexts. Achieves 60.47% on SWE-Bench Verified and 56.2% on SWE-Pro, competing with frontier closed models on agentic coding benchmarks. Supports 256K token context.

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.

Doubao Seed 2.0 Lite is ByteDance's general-purpose frontier model released February 14, 2026. Achieves 85.1 on GPQA Diamond and 87.7 on MMLU-Pro, matching or exceeding many larger models. Optimized for production workloads requiring strong general reasoning at moderate cost.

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.

ByteDance mid-tier Seed 2.0 agent model balancing strong performance with cost efficiency. Supports multimodal input (text, image, video), tool calling, and a 4-level reasoning effort system. AIME 2025: 93, Codeforces: 2233. Max output: 131K tokens.

Qwen3.6-35B-A3B is an open-weight multimodal MoE model with 35B total parameters and 3B activated per token, released April 2026. It features a hybrid architecture combining Gated DeltaNet linear attention and standard Gated Attention with 256 total experts (8 routed + 1 shared), and includes a vision encoder for image and video understanding. Optimized for agentic coding, long-context reasoning, and visual tasks; supports 256K native context (extensible to ~1M via YaRN) with integrated thinking mode for multi-turn agent interactions.

Claude Haiku 4.5 is Anthropic's Claude 4.5 model with multimodal text and image input. It offers a 200K-token context window and scores 73.3 on SWE-bench Verified.

Jais 3 MoVA 36B A4B is a member of Jais 3, the Arabic-native open-weight fleet from the Institute of Foundation Models (IFM, MBZUAI) and Inception42, built as an adaptation of IFM's K2 Horizon 36B (4B active) model. Released October 15, 2026 under Apache 2.0, it is a bilingual Arabic/English text model with reasoning and tool use and a 512K-token context window.

MAI-Code-1.1-Flash is Microsoft AI's successor to MAI-Code-1-Flash: a sparse Mixture-of-Experts coding model (138B total / 5B active per the model card) with a 256K context window, text and image input, and text output. It launched in GitHub Copilot on 2026-08-11 with native vision, stronger coding and tool use, and a lower list price than 1-Flash. Microsoft reports SWE-Bench Verified 72.6% and Terminal-Bench 2.1 62.9% in the Copilot harness, ~25% better token efficiency, and about one quarter the cost of the June 1.0 Flash launch. GitHub Copilot models-and-pricing (per 1M tokens): $0.20 input, $0.02 cached input, $1.20 output. On 2026-10-07 Microsoft added an on-device quantized build (Windows ML / Copilot local; recommend >120GB RAM) with zero per-inference charges for local calls. Proprietary product terms; not on Hugging Face.

Open-weight dense 27B multimodal Qwen3.5 model with native vision-language support. Apache 2.0.

K2-Horizon-7B is a medium dense K2 Horizon decoder-only model with 7B parameters, 512K native context, and Apache 2.0 weights on Hugging Face (IFM/K2-Horizon-7B). Released September 3, 2026 for agentic tool use, coding, long-context reasoning, and scientific benchmarks with k2_horizon reasoning and tool parsers.

Jais 3 32B is a member of Jais 3, the Arabic-native open-weight fleet from the Institute of Foundation Models (IFM, MBZUAI) and Inception42, built as an adaptation of IFM's K2 Horizon 32B model. Released October 15, 2026 under Apache 2.0, it is a bilingual Arabic/English text model with reasoning and tool use and a 512K-token context window.

Open-weight MoE Qwen3.5 model with 122B total and 10B active parameters. Apache 2.0.

OpenAI o3 reasoning model with advanced multi-step problem-solving capabilities.

GPT-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.

MAI-Code-1-Flash is Microsoft AI's lightweight agentic coding model built directly inside GitHub Copilot's production harness. It is designed for fast everyday developer workflows, adaptive thinking by task complexity, multi-turn instruction following, and token-efficient coding. Microsoft reports 51.2% on SWE-bench Pro versus 35.2% for Claude Haiku 4.5 in the same Copilot harness, plus stronger results on SWE-bench Verified, SWE-bench Multilingual, and Terminal-Bench 2.0 without publishing exact scores for those secondary benchmarks.

Qwen3-Coder-Next is an ultra-sparse Mixture-of-Experts coding agent model from Alibaba's Qwen team, released February 3, 2026 under Apache 2.0. It has 80B total parameters with 3B active at inference, delivering substantially higher throughput than comparable dense models. It supports a native 256K context window, function calling, structured outputs, Claude Code, Qwen Code, Cline, Kilo, and other scaffold templates. Reported benchmarks include SWE-Bench Pro 44.3%, SWE-Bench Resolved 70.6%, and TerminalBench 2 36.2%.

xAI's Grok Code Fast 1 is a code-specialized Mixture-of-Experts model released August 27, 2025.

NVIDIA's open frontier-reasoning model (550B total / 55B active MoE, hybrid Transformer-Mamba). Highest Artificial Analysis Intelligence Index for any US open model (score: 48). 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.

Claude 3.7 Sonnet is Anthropic's advanced model with extended thinking capabilities, offering state-of-the-art reasoning for complex tasks.

Jais 3 375B A23B is a member of Jais 3, the Arabic-native open-weight fleet from the Institute of Foundation Models (IFM, MBZUAI) and Inception42, built as an adaptation of IFM's K2 Horizon 375B (23B active) model. Released October 15, 2026 under Apache 2.0, it is a bilingual Arabic/English text model with reasoning and tool use and a 512K-token context window.

Jais 3 7B is a member of Jais 3, the Arabic-native open-weight fleet from the Institute of Foundation Models (IFM, MBZUAI) and Inception42, built as an adaptation of IFM's K2 Horizon 7B model. Released October 15, 2026 under Apache 2.0, it is a bilingual Arabic/English text model with reasoning and tool use and a 512K-token context window.

Alibaba's Qwen3.5-35B-A3B is a Mixture-of-Experts model released February 24, 2026, with 35B total parameters and 3B active during inference. Part of the Qwen3.5 series with a 262K native context window (extendable to ~1M tokens). Optimized for high inference throughput (78+ tokens/second on NVIDIA hardware). Open-source under Apache 2.0.

DeepSeek: DeepSeek V3.2 Exp available via OpenRouter. Pricing: $0.27/1M input, $0.41/1M output.

Doubao Seed 2.0 Mini is ByteDance's compact frontier model released February 14, 2026. Achieves 79.0 on GPQA Diamond and 83.6 on MMLU-Pro. Designed for high-throughput batch processing where cost and latency are critical.

A 30B-parameter sparse mixture-of-experts (3B active parameters) code generation model from Cohere, released as open weights under Apache 2.0. Designed for agentic software engineering, code generation, and terminal-based tasks. Supports a 256K token context window with 64K maximum output length, interleaved thinking, and tool use via JSON schema chat templates. The inaugural model in Cohere's North family.

Qwen3-Coder-480B-A35B-Instruct is Alibaba's flagship open-source code generation and agentic model, released July 22, 2025 under the Apache 2.0 license. The model has 480 billion total parameters with 35 billion active parameters per token, organized across 62 transformer layers with 160 specialized expert networks and 8 experts activated per token. It uses Grouped Query Attention (GQA) with 96 query heads and 8 key-value heads and supports a native context window of 262,144 tokens, extendable to 1 million tokens via YaRN position scaling. The model is purpose-built for software engineering tasks and agentic workflows: code generation, code review, test writing, multi-step debugging, and browser-based agentic task execution. On release, it achieved state-of-the-art results among open models on Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use benchmarks, with performance comparable to Claude Sonnet 4 on these tasks. Available via Fireworks AI, Google Vertex AI, NVIDIA NIM, AWS Bedrock, Novita AI, and the Vercel AI Gateway.

Kolibri (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.

Hybrid thinking/non-thinking text model from DeepSeek with tool use, agent support and a 128K context window.

K2-Horizon-3.7B is a small dense K2 Horizon decoder-only model: 3.7B parameters, 512K native context, and Apache 2.0 weights on Hugging Face (IFM/K2-Horizon-3.7B). Part of IFM's six-model open fleet alongside larger MoE and dense siblings.

Google DeepMind's most capable Gemini 2.5 model with native thinking/reasoning support. Features a 1M-token context window, multimodal inputs (text, image, audio, video), function calling, and strong performance across coding, mathematics, and scientific reasoning tasks.

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.

GLM-4.7 Flash is Tsinghua Knowledge Engineering Group (THUDM)'s GLM-4 model. It offers a 198K-token context window.

DeepSeek R1 0528 is DeepSeek's DeepSeek R1 model with an optional reasoning mode. It offers a 130K-token context window with weights openly available for self-hosting and scores 81 on GPQA.

IBM Granite 4.2 30B is the flagship of the Granite 4.2 family (released 25 August 2026): a dense decoder-only reasoning model (GraniteForCausalLM, ~29.3B params) post-trained from Granite-4.1-30B-Base with built-in <think> chain-of-thought, reasoning-augmented tool calling, 128K native context (512K long-context extension) and 12 tested languages. Apache 2.0.

MiniMax-M1-80k is MiniMax's MiniMax M1 model. It offers an 80K-token context window.

OpenAI's GPT-4.1 model released April 2025, excelling at coding tasks, precise instruction following, and web development. Outperforms GPT-4o in these areas with a 1 million token context window. Available via API and in ChatGPT for Plus, Pro, Team, Enterprise, and Edu users.

30B MoE model with 3B active parameters - superior reasoning with IMO/IOI 2025 gold-medal performance

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

Advanced reasoning and coding model with improved performance, supports cached context for cost savings

IBM Granite 4.2 8B (released 25 August 2026) is a dense decoder-only reasoning model (~8.8B params) post-trained from Granite-4.1-8B-Base with built-in <think> reasoning, tool calling, 128K native context (512K extension) and 12 tested languages. Apache 2.0; served on DeepInfra per HF provider mapping.

Mellum2.1 Thinking is JetBrains' 2026-10-08 open-weight upgrade of Mellum2: same 12B MoE / 2.5B active / 131K context architecture, with much heavier reinforcement learning in real repositories for agentic coding, tool use, math, and reasoning. Apache 2.0. JetBrains reports large gains on SWE-bench Verified (47.0 vs 2.0 for Mellum2 Thinking under the same Pi harness) and LiveCodeBench v6 (82.0). Serve with vLLM (reasoning-parser qwen3; optional Hermes tool calling). GGUF/MTP heads coming soon per the card. Open weights on Hugging Face; not listed as a hosted JetBrains API model.

Upstage Solar Pro 3 is Upstage's Solar Pro model. It offers a 200K-token context window.

Fast and efficient small model from OpenAI replacing GPT-4o mini. Released April 2025 alongside GPT-4.1. Shows improvements in instruction-following, coding, and intelligence with a 1 million token context window. Available in ChatGPT for paid users.