Models

32B dense model matching Qwen2.5-72B performance, supporting 119 languages with thinking mode toggle

2.4T sparse MoE (~95B active) flagship from Alibaba — text + image in, toggleable thinking with a 262K reasoning budget, agentic tool use, 1M context

2.4T sparse MoE (~95B active) flagship from Alibaba — text + image in, toggleable thinking with a 262K reasoning budget, agentic tool use, 1M context

671B MoE model (37B active) with hybrid thinking/non-thinking modes, tool use, and 128k context

284B MoE (13B active) fast tier of DeepSeek V4 with hybrid CSA/HCA sparse attention, toggleable deep reasoning, tool use, and 1M-token context

1.6T MoE model (49B active) with hybrid CSA/HCA sparse attention, toggleable thinking modes, tool use, and 1M context

552B multimodal MoE (8B active on prefill, 16B on decode) — DeepSeek's first Causal Encoder-Decoder model, pairing compressed sparse attention with FP4 KV caching for input-heavy agentic work: native text + image in, continuously controllable reasoning effort, tool use, and 1M context

Google Gemma 4 31B dense instruction-tuned multimodal model (text + image in) with a toggleable thinking mode, native tool use, 128k context, and 140+ language coverage

400B MoE (17B active, 128 experts) with native multimodal early fusion supporting 12 languages and images

70B instruction-tuned model delivering 405B-class text performance with 128k context and GQA

456B MoE model (27B active, 256 experts) from MiniMax with SOTA coding (80.2% SWE-Bench) and agentic tool use

230B sparse MoE (10B active, 256 experts, 8 per token) from MiniMax — agentic workflows, 200K context

428B MoE (23B active, 128 experts, 4 per token) from MiniMax — MiniMax Sparse Attention for 1M context, long-horizon agentic coding, and thinking that can be turned off; text + image in

Coding-focused agentic model from Moonshot AI built on Kimi K2.6 (DeepSeek-V3-style MLA, 384 routed experts plus one always-active expert) — always-on reasoning tuned for long-horizon software engineering, tool use, and 256K context

2.8T-parameter Stable LatentMoE (16 of 896 experts active) from Moonshot AI with Kimi Delta Attention — native text/image/video understanding, toggleable reasoning, tool use, and 1M-token context

117B open-weight MoE (5.1B active) from OpenAI achieving near o4-mini reasoning

High performance coding assistant with reasoning, optimized for algorithms, debugging, and code review

117B MoE model from scx.ai achieving near o4-mini reasoning

753B sparse MoE (~40B active) from Z.ai with IndexShare sparse attention — long-horizon agentic coding, toggleable High/Max thinking modes, 1M context

Latency-optimised preview of GLM-5.2 (753B sparse MoE, ~40B active) from Z.ai — same IndexShare sparse attention and toggleable thinking modes, tuned for faster time-to-first-token

743B sparse MoE from Z.ai on the GLM-5.2 base with scaled post-training — long-horizon agentic coding, always-on thinking with low/high/max effort, 1M context

Fast variant of Z.ai's GLM-5.3 with a 1M context window, 128K maximum output, always-on reasoning, and agentic tool use

Native multimodal model (text + image in) from Z.ai for efficient coding and long-horizon agent tasks — hybrid sparse and linear attention for accurate long-context behaviour, always-on thinking, 1M context