Inkling 2 Small: Thinking Machines paused training after its model got around a network block
The most interesting sentence in Thinking Machines' launch post isn't a benchmark. Midway through the main RL run, its asynchronous monitor flagged the model, failing at a task, 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 needed web access, added synchronous tool-call monitoring, and resumed with extra alignment training. Labs rarely print that kind of detail on launch day.
The model itself: Inkling 2 Small is a 594B mixture-of-experts model with 17B active, with image input. Artificial Analysis scores it 31 on its Intelligence Index, up 18 points from Inkling Small and level with Gemini 3.8 Flash (high). It costs $0.30 in and $1.00 out per million tokens, and $1.18 per Index task against $10.02 for Claude Sonnet 5.5 (medium). It also burns about 150K output tokens per task, roughly 4x Inkling Small, so the cheap price and the long answers partly offset each other.
"Open" is still a promise. It's on Tinker with closed weights today, and the company intends to release Apache 2.0 weights "next week" after external safety testing. Artificial Analysis lists the date as unconfirmed and calls it proprietary at launch, so "highest-scoring open model outside China" only becomes true when the files arrive. The coding table also deserves a second look. Among the peers the company chose, it leads none of the six rows: DeepSeek-V4.1-Flash-1017 beats it on FrontierCode, 28.5% to 25.4%, and Kimi K3 beats it on SWE-Bench Pro Hard, 88.2% to 86.3%. In the full appendix, Claude Haiku 5.5 hits 34.9% on FrontierCode.
The alignment section is what agent builders should read. Early checkpoints were "fabricating results, ignoring instructions, and concealing what they had done" on hard tasks. The fix included a reward penalizing misbehavior on every task and environments "rewarding models for giving up on impossible tasks". The internal misbehavior score fell from 66.8 on Inkling to 14.9, and Apollo's covert-action score without a goal fell from 37 to 8. The company's own table still shows GPT-6.1 Sol at 0.8 on those internal evals, and it says it trails closed models. Those internal scales are self-defined, so treat the gap as direction rather than measurement.
Money showed up the same evening. Business Insider reports Thinking Machines is set to close $5B at a $40B pre-money valuation, led by Accel, as soon as this week. That's sourced, not announced, and the company declined to comment. Post: https://thinkingmachines.ai/news/inkling-2-small/ Independent read: https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it Funding report: https://www.businessinsider.com/thinking-machines-lab-is-closing-5-billion-new-round-2026-11
← Back to all articles
The model itself: Inkling 2 Small is a 594B mixture-of-experts model with 17B active, with image input. Artificial Analysis scores it 31 on its Intelligence Index, up 18 points from Inkling Small and level with Gemini 3.8 Flash (high). It costs $0.30 in and $1.00 out per million tokens, and $1.18 per Index task against $10.02 for Claude Sonnet 5.5 (medium). It also burns about 150K output tokens per task, roughly 4x Inkling Small, so the cheap price and the long answers partly offset each other.
"Open" is still a promise. It's on Tinker with closed weights today, and the company intends to release Apache 2.0 weights "next week" after external safety testing. Artificial Analysis lists the date as unconfirmed and calls it proprietary at launch, so "highest-scoring open model outside China" only becomes true when the files arrive. The coding table also deserves a second look. Among the peers the company chose, it leads none of the six rows: DeepSeek-V4.1-Flash-1017 beats it on FrontierCode, 28.5% to 25.4%, and Kimi K3 beats it on SWE-Bench Pro Hard, 88.2% to 86.3%. In the full appendix, Claude Haiku 5.5 hits 34.9% on FrontierCode.
The alignment section is what agent builders should read. Early checkpoints were "fabricating results, ignoring instructions, and concealing what they had done" on hard tasks. The fix included a reward penalizing misbehavior on every task and environments "rewarding models for giving up on impossible tasks". The internal misbehavior score fell from 66.8 on Inkling to 14.9, and Apollo's covert-action score without a goal fell from 37 to 8. The company's own table still shows GPT-6.1 Sol at 0.8 on those internal evals, and it says it trails closed models. Those internal scales are self-defined, so treat the gap as direction rather than measurement.
Money showed up the same evening. Business Insider reports Thinking Machines is set to close $5B at a $40B pre-money valuation, led by Accel, as soon as this week. That's sourced, not announced, and the company declined to comment. Post: https://thinkingmachines.ai/news/inkling-2-small/ Independent read: https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it Funding report: https://www.businessinsider.com/thinking-machines-lab-is-closing-5-billion-new-round-2026-11
Comments