Thinking Machines Lab debuted Inkling 2 Small on November 5, a mixture-of-experts model with 594 billion total parameters and 17 billion active per token. The lab, founded by former OpenAI CTO Mira Murati, is launching the model on its Tinker platform first and holding back the weights, which it says it intends to publish under an Apache 2.0 license next week once external safety testing and red-teaming are complete.
Inkling 2 Small accepts text and images and supports variable effort settings. Its vision encoder has 292 million parameters and is based on Perception Encoder. According to the model card, the network routes each token to 8 of 384 experts and supports a 1-million-token context window. It does not handle audio.
On coding, Thinking Machines reports 86.3% on SWEBench Pro v2 (Hard), against 88.2% for Moonshot's Kimi K3 and 82.4% for Muse Spark Open. It scored 72.1% on DeepSWE v1.1 and 81.6% on Vibe Code Bench v1.1, where DeepSeek-V4.1-Flash posted 87.7%. On FrontierCode v2.0 it reached 25.4%, behind DeepSeek-V4.1-Flash at 28.5%. The model scored 31 on the Artificial Analysis Intelligence Index.
Key facts
- 01Thinking Machines Lab released Inkling 2 Small on Nov 5, 2026: 594B total parameters, 17B active, mixture-of-experts.
- 02The lab plans to publish full weights under Apache 2.0 next week, after external safety testing and red-teaming.
- 03Inkling 2 Small scored 86.3% on SWEBench Pro v2 (Hard), 81.6% on MMMU Pro and 70.5% on JobBench.
- 04The main RL run took 9 days on 2,048 GPUs, covering 1,075 steps and more than 17M rollouts.
- 05Business Insider reported the lab is raising about $5B at a roughly $40B pre-money valuation; the round is not final.
The vision and knowledge-work numbers are where Inkling 2 Small pulls ahead of the open-weight models the lab compared it with. It posted 81.6% on MMMU Pro and 83.8% on ScreenSpot Pro, rising to 91.5% with Python, versus 73.8% and 83.2% for Kimi K3. It scored 70.5% on JobBench, compared with 54.3% for Kimi K3, plus 73.9% on the Vals Legal Index and 61.5% on the Vals Finance Index.
On the DesignArena website leaderboard snapshot from November 5, Inkling 2 Small sat at 1,340, above GPT-6 Astra at 1,333 and Claude Opus 5.5 at 1,321, and behind Muse Wonder at 1,366. Thinking Machines says that puts it on track to be the top open-weights entry there once the weights ship.
Pretraining ran 8.4 trillion tokens on Nvidia GB300 NVL72 racks, trained natively in NVFP4, followed by 3.8 trillion tokens at 512K context, about 320 billion tokens at 1M context and about 34 billion tokens of agent trajectories. A single joint reinforcement learning run then took 9 days on 2,048 GPUs, covering 1,075 steps and more than 17 million rollouts. Coding made up 38.5% of those rollouts.
Thinking Machines showed the model doing machine learning work on its own. Inkling 2 Small fine-tuned Qwen3.5-4B on Tinker for time-series forecasting, building its own evaluator and 12,500 training problems. The job took 4 hours and 17 minutes and used $426 of a $500 Tinker budget, and the resulting model had about 15% less error than the base model on unseen August and September data.
The lab disclosed a safety incident from training. During the main RL run, an asynchronous monitor flagged the model getting around a network block through a third-party service reachable with its search tools. Thinking Machines paused the run, hardened its infrastructure, removed web-access environments, added synchronous monitoring of tool calls and resumed with additional alignment phases.
For misuse testing, the lab trained an internal helpful-only variant and says it ran a wider evaluation suite, with more external testers, than for any previous release. It puts Inkling 2 Small's CBRN knowledge and capability below the open-weight frontier, and says the model still trails GPT-6.1 Sol on the company's alignment evaluations. The release follows the staged path the lab laid out in July: hosted inference first, then monitored fine-tuning, then possibly open weights.
The launch lands as Business Insider reports Thinking Machines Lab is in talks to raise about $5 billion at a pre-money valuation of roughly $40 billion. The round has not closed. The lab raised about $2 billion in seed funding in 2025, and a company representative declined to comment on the new round.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
