The short version
● Size — 594B total, 17B active, sparse MoE
● Context — 1M tokens, with compaction for long agent runs
● Coding — 86.3% SWE-Bench Pro v2 Hard, 72.1% DeepSWE v1.1
● Weights — closed at launch, Apache 2.0 promised the week after
Thinking Machines announced Inkling 2 Small on 5 November. It is a sparse mixture-of-experts transformer with 594B total parameters and 17B active, paired with a 292M-parameter vision encoder based on Perception Encoder. Context runs to 1M tokens with compaction for long agent trajectories, it takes text and image input and returns text, there is no audio, and reasoning effort is variable.
The reported scores: 86.3% on SWE-Bench Pro v2 Hard, 72.1% on DeepSWE v1.1, and 31 on the Artificial Analysis Intelligence Index. The number the lab seems proudest of is an internal misbehaviour rate of 14.9 against 66.8 for the first Inkling, where lower is better — a four-fold reduction, on a metric only they can measure. Training used NVFP4 pretraining on NVIDIA GB300 NVL72 racks and a nine-day reinforcement-learning run across 2,048 GPUs.
Availability is the interesting part. It is API-only through Tinker, with weights closed at launch under the Tinker terms, and the lab says it intends to release full weights under Apache 2.0 the following week once external safety testing and red-teaming finish. No exact date is confirmed and no pricing has been published. Announcing an open-weight release a week before it happens is unusual — it is a commitment that can be checked, which is more than most open-weight promises offer. Specs are on the Inkling 2 Small listing, and our comparison against Hunyuan Hy4 puts the two large MoE releases of this week side by side.
More from 7 November
Full dated index: every AI release in November 2026.
Sources