594B total / 17B active (sparse MoE; 8 of 384 experts per token)
1,048,576 tokens
Apache 2.0
Open-weight general-purpose multimodal model accepting text and image inputs and generating text, with particular emphasis on agentic coding, tool use, knowledge work, visual reasoning, long-context operation, retrieval-augmented generation, and conversational applications. Its 48-layer decoder-only transformer uses a sparse Mixture-of-Experts feed-forward backbone with each token routed to eight of 384 experts, hybrid local and global attention, and a native 23-layer vision transformer that projects images into the decoder's shared hidden space. The text backbone has a 4,608-dimensional hidden state, 48 attention heads, eight key-value heads, 2,304-dimensional expert feed-forward layers, and a 201,024-token padded vocabulary. Inkling 2 Small supports structured tool calling and continuously variable reasoning effort from 0.0 to 0.99, with named presets from none through max and a default effort of 0.9. Training began with 8.4 trillion natural-language, code, and image tokens in a 68-hour NVFP4 run on NVIDIA GB300 NVL72 systems. Later phases added 3.8 trillion higher-quality and newer tokens at 512K context, approximately 320 billion reasoning and agent-trajectory tokens at one-million-token context, and approximately 34 billion highest-quality agent-trajectory tokens. Its unified reinforcement-learning run used 2,048 GPUs for nine days, exceeded 17 million rollouts and three trillion generated tokens, and covered coding, office work, STEM, web and app interaction, vision, tool use, search, general agents, chat, safety, and epistemics. Coding training included hundreds of thousands of long-horizon tasks across multiple agent harnesses and context-compaction strategies; early checkpoints were tested through OpenCode under the stealth name Fledge-Alpha. Publisher-reported evaluations, generally run at 0.99 effort, temperature 1.0, and a one-million-token maximum trajectory, include 30.6 on Artificial Analysis Intelligence Index v5.0, 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 Hard, 81.6% on Vibe Code Bench v1.1, 90.2% on BrowseComp with context management, 85.4% on MCP Atlas Full, 1,527 Elo on AA-Briefcase v1.1, 70.5% on JobBench, and 83.3% on AA-LCR v1.1. The model was pretrained, midtrained, and post-trained natively in NVFP4. Thinking Machines recommends the approximately 393.5 GB NVFP4 checkpoint, distributed as 24 model shards plus a separate approximately 3.42 GB MTP file; it requires at least 500 GB of aggregate VRAM, such as two B300 or four H200 GPUs. The approximately 622.8 GB MXFP8 checkpoint uses 48 model shards plus the same MTP file, requires at least 800 GB of aggregate VRAM, and reportedly gains only one to two percentage points across evaluations. Documented local inference paths include SGLang, vLLM, TokenSpeed, Unsloth, and Transformers. The publisher notes that the model can be verbose, may choose suboptimal tools in repositories containing many skills, is less polished for ordinary chat, and does not support audio.