# Language Models ## Commercial - **Anthropic Claude** — [claude.ai](https://claude.ai) - Fable 5 (newest frontier model) - Opus 4.8 (highest capability; complex reasoning and agentic tasks) - Sonnet 4.6 (balanced intelligence and speed; production workhorse) - Haiku 4.5 (fast; lowest cost) - **OpenAI GPT** — [openai.com](https://openai.com) - GPT-5.5 (latest flagship) - GPT-5.4 (production workhorse; 1M context) - o3 / o4-mini (reasoning-focused; multi-step math, code, science) - o3-pro (highest reasoning quality, highest cost) - **Google Gemini** — [gemini.google.com](https://gemini.google.com) - Gemini 3 Pro (frontier; 1501 Elo LMArena, GPQA Diamond 91.9%) - Gemini 3.5 Flash (near-Pro intelligence at Flash cost; parallel agentic execution) - Gemini 2.5 Pro (1M context, strong reasoning and coding) - Gemini 2.5 Flash (fast; controllable thinking budget) ## Open for commercial use - **DeepSeek V4** (MIT) — [deepseek.com](https://www.deepseek.com) - V4 Pro: 1.6T parameter Mixture-of-Experts (MoE), 49B active parameters, 1M context window - V4 Flash: 284B total / 13B active, cost-efficient variant - [DeepSeek V4 Pro overview](https://deepinfra.com/blog/deepseek-v4-pro-model-overview) - **GLM-5** (ZhipuAI / Tsinghua, MIT) — [zhipuai.cn](https://www.zhipuai.cn) - 744B MoE; 1M context; leads Intelligence Index and SWE-bench Pro (58.4%) - **Qwen3** (Alibaba, Apache 2.0) — [qwen.ai](https://qwen.ai) - Qwen3.5-397B-A17B: 397B total / 17B active MoE flagship, 200+ language support - Qwen3.6-27B: 27B dense, 256K context (extensible to 1M); beats 397B on coding (SWE-bench 77.2%) - Qwen3.6-35B-A3B: 35B total / 3B active MoE; 3–4× faster than 27B with moderate quality loss - **MiniMax M3** (MiniMax, open-weight) — [minimaxi.com](https://www.minimaxi.com) - 428B total / ~23B active MoE; 1M context, native image and video input; SWE-bench Pro 59.0% - **Gemma 4** (Google, Apache 2.0) — [ai.google.dev](https://ai.google.dev) - E2B / E4B: 2B / 4B effective params, edge/mobile deployment, 128K context - 12B: dense, unified text/image/audio, 128K context - 31B: dense, server-grade, 256K context - 26B A4B: 26B total / 4B active MoE, high-throughput reasoning, 256K context - **Llama 4** (Meta, custom license) — [llama.com](https://www.llama.com/models/llama-4/) - Scout 17B-16E: 109B total / 17B active, 16 experts, 10M context; fits single H100 - Maverick 17B-128E: 400B total / 17B active, 128 experts, 1M context - Behemoth 288B-16E: ~2T total / 288B active, 16 experts (preview) - **Hy-MT2** (Tencent, Apache 2.0) — [github.com/Tencent-Hunyuan/Hy-MT2](https://github.com/Tencent-Hunyuan/Hy-MT2) - Translation-specialized family (33 languages, fast-thinking mode) - 1.8B / 7B: dense variants; 1.8B fits on-device (440 MB at 1.25-bit quant) - 30B-A3B: 30B total / 3B active MoE; outperforms DeepSeek V4 Pro on translation ## Leaderboards - [Artificial Analysis LLM Leaderboard](https://artificialanalysis.ai/leaderboards/models) - [Vellum LLM Leaderboard](https://www.vellum.ai/llm-leaderboard) - [CodeSOTA Coding Benchmark](https://www.codesota.com/llm)