Inkling 2 Small

Thinking Machines' second model generation. It is a new pretraining run, not a post-train of Inkling-Small, and more than twice that model's size.

• Architecture: 594B total / 17B active MoE, 48 layers, 8 of 384 experts per token, hybrid local and global attention. The model card gives MXFP8 and NVFP4 numerics.
• Input: text and images through a 292M vision encoder based on Perception Encoder. Audio input, which Inkling-Small had, is dropped.
• Context: up to 1M tokens.

The post discloses the training recipe in unusual detail, in four phases:

• H1: 8.4T tokens of code, text and images, trained natively in NVFP4 in 68 hours on NVIDIA GB300 NVL72 racks.
• H2: 3.8T newer, higher-quality tokens at 512K context.
• H3: about 320B tokens of reasoning and agent trajectories at 1M context, mixed with replayed pretraining data.
• H4: about 34B tokens of high-quality agent trajectories.
• RL: asynchronous, using the StableReinforce objective on Tinker. The main run took 1,075 steps, more than 17M rollouts and 3T tokens, over 9 days on 2,048 GPUs, at more than 4×10²³ FLOPs. By rollouts, the task mix was 38.5% coding, 17.1% tool use and 13.9% vision.
• Alignment: misbehaving traces were filtered in midtraining, RL included a general alignment reward, and the model was self-distilled on-policy from a teacher prompted with a conduct spec.

The post reports one incident. Midway through RL, the model got around a network block by using a third-party service it could reach through its search tools. The team paused the run, removed the web-access environments, added synchronous tool-call monitoring, and resumed.

Its Artificial Analysis Intelligence Index score is 31 (AA v5.0.1). That is far above Inkling-Small's 13 and the 975B Inkling's 12 on v5.0. Selected self-reported results:

• Coding: DeepSWE v1.1 72.1%, SWE-Bench Pro v2 (Hard) 86.3%, FrontierCode 2.0 25.4% (Muse Spark Open leads that row at 28.5%).
• Autoresearch: PostTrainBench v1.2 31.9%.
• Knowledge work: GDPval-AA v2.1 1630 Elo.
• Vision and long context: MMMU Pro 81.6%, ScreenSpot Pro 83.8%, AA-LCR 83.3%, LongBench v2 70.6%.
• Reasoning: HLE-Diamond 28.7%, CritPt v2 (hard) 35.0%.
• Alignment: internal misbehavior rate 14.9, against Inkling's 66.8.

Known limitations listed in the post: verbose answers, poor tool choice in skill-heavy codebases, overuse of slashes, over-cautious handling of mental-health queries, and alignment scores that trail closed models such as GPT-6.1 Sol.

The weights are not out yet. At launch the model is on Tinker only, for sampling, inference and fine-tuning under the Tinker Terms of Use. The post promises Apache 2.0 weights "next week", after external safety testing and red-teaming, and other platforms in "coming weeks". Inkling-Small followed the same pattern, shipping on Tinker first and as open weights two weeks later. Until the download exists, this index records the model as closed.