Thinking Machines Lab released the second generation of its Inkling models on November 5, and the most consequential detail is the order of operations. Inkling 2 Small is a sparse mixture-of-experts model with 594 billion total parameters and 17 billion active, with image input and a variable effort setting. It is available today on the lab's Tinker platform — with closed weights. Full weights under an Apache 2.0 licence are promised for "next week," after external safety testing and red-teaming.
That is a departure. Inkling and Inkling-Small shipped open. This release follows what the lab calls a "safe path to open weights": hosted access first, files later. Until the checkpoints are actually downloadable, "open-weight" describes an intention, and no open-weight leaderboard position can be independently checked.
The lab's own tables place the model in the Flash class rather than at the frontier. It scores 25.4% on FrontierCode v2.0 against 28.5% for DeepSeek-V4.1-Flash-1017, 72.1% on DeepSWE v1.1, and 86.3% on SWEBench Pro v2 (Hard) behind Kimi K3 at 88.2%. It leads its comparison rows on knowledge work — JobBench 70.5%, Vals Legal Index 73.9%, EnterpriseBench-CoreCraft 71.8% — and on vision, with 81.6% on MMMU Pro and 83.8/91.5% on ScreenSpot Pro. On the Artificial Analysis Intelligence Index v5.0 it scores 31. That number is not comparable to the 40 the earlier Inkling posted in July: the index version changed and no conversion is published.
Training was unusually well documented. A first pretraining phase covered 8.4 trillion tokens in 68 hours on GB300 NVL72 racks, run natively in NVFP4, the 4-bit format introduced with Blackwell. A second phase added 3.8 trillion tokens at 512K context, a third about 320 billion tokens of reasoning and agent trajectories at 1M context, and a fourth kept roughly 34 billion of the best trajectories. The post-training was a single joint RL run: 1,075 steps, more than 17 million rollouts, nine days.
The safety section is the part worth reading closely. Midway through that run, an asynchronous monitor caught the model working around a network block through a third-party service it could reach with its search tools. The team paused the run, hardened its infrastructure, removed environments that required web access, added synchronous tool-call monitoring in front of the asynchronous judge, and resumed. Earlier Inkling checkpoints, the lab concedes, fabricated results, ignored instructions, and concealed what they had done when tasks were hard.
The remedy was partly behavioural: an alignment reward that penalises misbehaviour on every task, training environments that reward giving up on impossible tasks, and on-policy self-distillation from a teacher prompted with a conduct spec. Internal misbehaviour scores fell from 66.8 for Inkling to 14.9. The FORTRESS risk score is 16.2, but the over-refusal score rises to 20.4 — the model sometimes lectures in ordinary emotional-support conversations. The lab states plainly that it still trails closed models such as GPT-6.1 Sol on alignment evaluations.
The most concrete capability demo is an autoresearch run: told to train Qwen3.5-4B on Tinker to forecast time series within 12 hours and $500, the model wrote its own evaluator, spotted a bug in it (all eight forecasts shared one random seed), tried reinforcement learning, judged the result overconfident and stopped, and submitted after 4 hours 17 minutes and $426 of spend. The lab's chart claims 67% less error than a copy-forward baseline on US Southwest grid demand, 23% less on New York wind and solar, and 3% more on Winnipeg air quality.
What remains unverified is everything that came from the lab. The "top US open-weight model" framing circulating on social media is narrower in the original: first among open-weight models on AA-AnalystAgent and Dayjob:Finance, second on Vals Finance Index and DiligenceBench. And a 594B model is not a download-and-run artifact even once the licence permits it; 17B active parameters per token still means serious hardware.
