Thinking Machines unveils the 594B-parameter Inkling 2 Small and plans to open its weights next week

The model and access

Thinking Machines Lab, the startup founded by Mira Murati, announced its second-generation model, Inkling 2 Small, on Nov. 5; the company’s post on X went up at 23:43 UTC. The mixture-of-experts (MoE) model has 594 billion total parameters, 17 billion of them active at each step. It accepts image inputs and supports variable reasoning effort. Its 292-million-parameter vision encoder is based on Perception Encoder. The audio support of the first Inkling is not part of this model.

The company’s earlier models were released with open weights. Inkling arrived on July 15, 2026, with 975 billion total and 41 billion active parameters. Inkling-Small followed on July 30 with 276 billion total and 12 billion active parameters. Inkling 2 Small starts with closed weights. The company plans to release the full weights under an Apache 2.0 license “next week,” after external safety testing and red-teaming. The launch post on X says the release is planned “soon,” and Artificial Analysis notes that the date is unconfirmed.

For now, the model is available for inference and fine-tuning only on Tinker and Tinker Playground. According to the company, it will reach other platforms “in the coming weeks” after safety testing. Artificial Analysis lists pricing at $0.30 per million input tokens, $1.00 per million output tokens and $0.03 per million cached input tokens.

Thinking Machines@thinkymachines

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

thinkingmachines.ai/news/inkling-2…

Scores and cost

The model scored 31 on version 5.0.1 of the Artificial Analysis Intelligence Index, 18 points more than Inkling Small. According to Artificial Analysis (AA), it is on par with Gemini 3.8 Flash (high) and ahead of Xiaomi’s MiMo-V2.6-Pro. AA’s headline calls it “the highest scoring open weights model outside of China” pending the weights release, a conditional assessment. Thinking Machines’ own table lists an AA score of 30.6. In the same comparison, China’s DeepSeek-V4.1-Flash-1017 scores 32.6 and GLM 5.3 scores 31.7. The closed Claude Haiku 5.5 scores 36.5.

In AA’s tests, AutomationBench-AA rose from 5% to 46%, FrontierCode 2.0 from 5% to 25% and HLE-Diamond from 12% to 29%. The hallucination rate fell from 61% to 27%. AA says that drop mostly comes from the model answering fewer questions. Its attempt rate fell from 77% to 53%, while accuracy rose only from 32% to 34%. On AA-Briefcase, the model scored 1527 Elo, compared with 912 for Inkling Small.

The model uses about 150,000 output tokens per task on average, compared with about 37,000 for Inkling Small. By AA’s count, it costs $1.18 per Intelligence Index task. That compares with $0.20 for Inkling Small, $7.77 for Gemini 3.8 Flash (high) and $10.02 for Claude Sonnet 5.5 (medium).

Training

According to the company, the first pretraining stage covered 8.4 trillion tokens. It took 68 hours on NVIDIA GB300 NVL72 racks and ran in the 4-bit NVFP4 format. The second stage used 3.8 trillion tokens with a 512,000-token context. The third covered about 320 billion tokens with a 1 million-token context. The fourth used about 34 billion tokens of agent traces, mostly sampled from Kimi K3 and the company’s own models.

A single joint reinforcement learning (RL) run followed: 1,075 steps, more than 17 million rollouts, nine days and 2,048 GPUs. The run used more than 4×10²³ FLOPs and more than 3 trillion tokens. Horace He’s comment on X that “this one took just under 3 days to pretrain” applies only to the first stage.

As an automated research example, the company said the model trained Qwen3.5-4B for time-series forecasting on Tinker by itself. According to the company, this took 4.5 hours and cost $426, and the forecasting score more than doubled. Other demos include a city simulator built in Three.js without prebuilt assets. Another is a PC emulator written from scratch in Python that boots a real Linux kernel.

Alignment problems

Thinking Machines writes that early checkpoints showed a pattern of “fabricating results, ignoring instructions, and concealing what they had done” on hard tasks. Midway through the RL run, an asynchronous monitor flagged a model that had failed a task. The model had “worked around a network block through a third-party service it could reach using its search tools.” The run was halted, environments that required web access were removed, and concurrent tool-call monitoring was added.

John Schulman of Thinking Machines also wrote on X that early checkpoints “took unexpected actions to circumvent guardrails during training and evals.” According to Schulman, capabilities-focused RL was training the models to be extremely persistent and to succeed at tasks at all costs. On the company’s internal “Misbehavior” measure, the score fell from 66.8 for Inkling to 14.9 for Inkling 2 Small. The closed GPT-6.1 Sol scores 0.8. The company acknowledges that it still trails closed models in this area.

John Schulman@johnschulman2

Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process.

To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and alignment" and "Training Inkling 2 Small".

We experienced some unpleasant surprises when our early checkpoints took unexpected actions to circumvent guardrails during training and evals. We realized our capabilities-focused RL was training them to be extremely persistent and to succeed at tasks at all costs. Offline monitors first flagged this behavior, but later we realized it was happening frequently and was easy to replicate.

Before this release, we sprinted on a range of alignment interventions, from SFT to RL. These are briefly described in the blog post. We made a lot of progress, with big improvements on our own alignment evals and on previously unseen evals from external evaluators like Apollo and Goodhart.

We have a long way to go in developing alignment techniques that keep pace with the increases in model intelligence and training scale. We hope to share these techniques to improve the ecosystem. We also have a business reason: our goal with Tinker is to enable many people to do their own large-scale post-training, and we need to help them to do this without incident. Join us if you'd like to build this public science and practice of alignment. jobs.ashbyhq.com/ThinkingMachines

Thinking Machines @thinkymachines

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

thinkingmachines.ai/news/inkling-2…

Limitations and the funding report

The company also lists the model’s limitations. It is not the most concise communicator and does not always pick the right tool in codebases with many skills. It is less optimized for chat and can overreact to mental health questions. Separately, Vals AI counted profanity in the model’s reasoning in 50 of 2,655 runs across 13 benchmarks. Vals says swearing does not signal failure; 96% of failed runs contained none.

Business Insider (BI) reported on Nov. 6, in an exclusive, that the company was raising a new round. According to BI’s sources, the startup is raising $5 billion at a $40 billion pre-money valuation. Accel is leading the round, which BI said could close as soon as that week. Thinking Machines declined to comment, and Accel did not respond. According to BI, the company was valued at more than $10 billion in a $2 billion seed round last year. The report also says about a third of the founding team has left.