Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, has released Inkling 2 Small, its second-generation AI model. It is a 594-billion-parameter mixture-of-experts model that activates only about 17 billion parameters per request, accepts images as well as text, and is built for coding and long-running agent work. It is available now through the company’s Tinker platform, and Thinking Machines says the full weights will follow under an Apache 2.0 license once outside safety testing is finished.
Key takeaways
- Inkling 2 Small has 594B total parameters, with roughly 17B active per request, plus image input through a separate vision encoder.
- It scores 31 on the Artificial Analysis Intelligence Index, up 18 points from the first Inkling Small.
- Access today is through Tinker (inference, fine-tuning and a web Playground) with closed weights; Apache 2.0 weights are promised after external red-teaming.
- It is strong on agentic coding and vision tasks but uses far more tokens per task than its predecessor.
- The launch lands alongside a reported $5B funding round, which the company has not confirmed.
What happened
Thinking Machines announced the model on November 5 in a long technical post, Introducing Inkling 2 Small. The company describes it as the first release of its second model generation, with more variants in the Inkling 2 family planned.
The headline specs are a mixture-of-experts (MoE) design with 594 billion total parameters, of which about 17 billion are used for any given token. That split matters: MoE models store a lot of knowledge across many “expert” sub-networks but only run a small slice at a time, which keeps inference costs closer to those of a much smaller dense model. Inkling 2 Small also takes image inputs through a vision encoder of about 292 million parameters, and it supports “variable effort”, meaning callers can trade speed for more thorough reasoning.
On training, the company says pretraining began with 8.4 trillion tokens of text, code and image data, processed natively in NVIDIA’s low-precision NVFP4 format on GB300 NVL72 racks in 68 hours. Unlike the first Inkling, reasoning and agentic skills were trained together in a single reinforcement-learning run, which lasted nine days on 2,048 GPUs and completed more than 17 million rollouts, according to both the official post and Techstrong.ai’s report.
How Inkling 2 Small performs
Independent benchmarker Artificial Analysis gives Inkling 2 Small a score of 31 on its Intelligence Index, an 18-point jump over the original Inkling Small. The Decoder notes that puts it roughly level with Google’s Gemini 3.8 Flash and at the top among Western open-weight models, though some Chinese open models still score higher on several tests.
A few numbers from the company’s tables and Artificial Analysis stand out for builders:
- Agentic coding: 72.1% on DeepSWE and 86.3% on the hard split of SWE-Bench Pro v2. Kimi K3 edges it on the latter at 88.2%, and DeepSeek V4.1 Flash leads several other coding tests.
- Automation: AutomationBench-AA rose from 5% with the first Inkling to 46%, and FrontierCode 2.0 from 5% to 25%.
- Vision: 81.6% on MMMU Pro and the best score among the compared models on GDP.pdf, a document question-answering test.
- Hallucinations: the hallucination rate fell from 61% to 27%, but the model also declines to answer more often.
It is not a clean sweep. On Terminal Bench Science, a newer agentic test, Inkling 2 Small scored 5.7%, behind models such as Claude Haiku 5.5 and DeepSeek V4.1 Flash. Thinking Machines itself lists weaknesses: the model sometimes communicates inefficiently, does not always pick the best tools in large codebases, has no audio support and is less tuned for casual chat.
Why it matters
For most of the past year, the strongest open-weight models have come from Chinese labs such as DeepSeek, Moonshot (Kimi), Zhipu (GLM) and MiniMax. Western open releases have been fewer and, in some cases, carry restrictive licenses. Inkling 2 Small is notable because it pairs near-frontier agentic scores with a promised Apache 2.0 license, one of the most permissive options available, allowing commercial use and modification without the custom terms attached to some rival models. We saw a similar push last week when Mistral Large 4 went open-weight, though with a less permissive license.
The release also shows where Thinking Machines is placing its bet. Its first product, Tinker, is a platform for fine-tuning and sampling models, and the company says Inkling models are built to be strong base models for customization. Rather than competing head-on with ChatGPT or Claude as consumer assistants, it is courting teams that want to train their own specialised agents.
What it means for developers and businesses
You can try it now, but you can’t self-host it yet. Today the model runs only on Tinker, where developers can sample it, fine-tune it and test it in the Playground. Downloadable checkpoints are tied to the weights release, which the company says will come after external red-teaming. Until those files are public, treat the open-weight promise as a plan, not a fact.
Watch the token bill. According to Artificial Analysis figures cited by The Decoder and Techstrong, Inkling 2 Small uses about four times as many output tokens per task as its predecessor, and its average cost per Intelligence Index task is about $1.18 versus $0.20 for the first Inkling Small. That is still far below the $7.77 reported for Gemini 3.8 Flash and $10.02 for Claude Sonnet 5.5 on the same measure, but it means longer, more expensive runs if you leave effort set high.
Self-hosting will need serious hardware. Only 17 billion parameters are active at once, but all 594 billion still need to be loaded into memory. Expect multi-GPU server setups, not a laptop, once the weights arrive. Smaller teams will likely rely on hosted inference first.
It fits agent and UI-generation workloads. The company highlights tool calling over MCP and using visual feedback to build and refine websites, games and 3D designs. Its 1,340 score on DesignArena’s website test is the highest among open-weight models in its comparison. If you are building coding agents, document-processing pipelines or design tools and want a model you can eventually fine-tune and run privately, it is worth adding to your evaluation set.
The funding backdrop
The launch comes as Business Insider reports that Thinking Machines is close to raising about $5 billion in a round led by Accel at a $40 billion pre-money valuation, with Nvidia separately in talks to invest. The company declined to comment, so those figures remain unconfirmed. For context, Thinking Machines raised a $2 billion seed round at a roughly $12 billion valuation in July 2025, and in March announced plans with Nvidia to deploy at least one gigawatt of Vera Rubin systems starting in 2027. The AI funding climate has been shaky this week, as we covered in our look at the recent tech selloff, which makes a raise of this size a notable vote of confidence if it closes.
What to watch next
- The weights drop: whether the Apache 2.0 checkpoints land on schedule, and whether any usage conditions are added.
- Independent testing: results from the community once people can run the model outside Tinker, especially on real codebases.
- Hosting partners: which inference providers add it, and at what per-token price.
- Larger Inkling 2 models: the “Small” label implies bigger siblings are coming.
- The funding round: official confirmation of the reported $5 billion raise.
Inkling 2 Small is not the best model at everything, and its open-weight status is still a promise. But a 594B model with strong agentic scores, image input and a permissive license on the way gives developers a serious new Western option in open AI, and one built to be fine-tuned.