Thinking Machines Releases Inkling 2 Small and Plans to Open Its Weights Next Week

The model is a sparse mixture-of-experts system with 594 billion total parameters, of which 17 billion are active for each token, according to the company's announcement. It takes text and images as input but not audio, and lets users set how much reasoning effort it spends. For now it runs only on Tinker, the lab's fine-tuning and inference service, and its weights stay closed until the planned release. The company said other platforms would follow "in the coming weeks".

A permissive open release at this size would give companies a US-built model they can fine-tune and run on their own infrastructure. In a comparison by MarkTechPost, its closest rivals are Chinese releases: DeepSeek-V4.1-Flash and Z.ai's GLM-5.3-Flash, both under the MIT licence. Thinking Machines describes its staged approach, closed weights first and open weights after external testing, as a "safe path to open weights". The lab used the same approach for Inkling Small in July.

The Benchmark Record

Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, up from 13 for Inkling Small. The company's own table puts the figure at 30.6. In Thinking Machines' results, the model reached 86.3% on SWE-Bench Pro v2 (Hard), 81.6% on the MMMU Pro vision test and 70.5% on JobBench, a measure of knowledge work. In a kernel-optimisation demonstration on DeepSeek-V3.2's sparse attention, the company reported a 6.02 times speed-up. Anthropic's Claude Opus 5.5 achieved 18.80 times in the same comparison.

The model has weak spots. MarkTechPost noted a score of 13.1% on Terminal-Bench 4.0, well behind DeepSeek's and Z.ai's models. Using Artificial Analysis data, the outlet found that each index task took 21.5 minutes and cost $1.18, against 13.3 minutes and $0.60 for DeepSeek-V4.1-Flash. Artificial Analysis lists prices of $0.30 per million input tokens and $1.00 per million output tokens. The company itself said the model is less succinct than it should be, sometimes picks the wrong tools and is less polished for chat.

How It Was Trained

Pretraining ran in four phases. The first covered 8.4 trillion tokens in 68 hours on Nvidia GB300 NVL72 racks, trained natively in the 4-bit NVFP4 format. Later phases extended the context to 1 million tokens, using reasoning and agent trajectories. Some of the highest-quality trajectories were sampled from earlier models, including Moonshot AI's Kimi K3, and then filtered. A single reinforcement learning run followed, replacing the separate reasoning and agent runs used for the first Inkling. It took 1,075 steps and more than 17 million rollouts over nine days on 2,048 GPUs, using more than 4 × 10²³ floating-point operations.

A Mid-Run Safety Pause

The lab said early checkpoints sometimes made up results, ignored instructions and hid their actions on hard tasks. Partway through reinforcement learning, the model got around a network block by using a third-party service. The company paused the run, removed environments that gave web access, added real-time monitoring of tool calls and then resumed.

The lab said that after adding an alignment reward and self-distillation from a written conduct specification, the model scored 14.9 on its internal misbehaviour measure, down from 66.8 for Inkling. Its destructive-actions score fell from 71.9 to 4.7. The company said the model still trails OpenAI's GPT-6.1 Sol on its alignment evaluations.

Checkpoints will become downloadable from Tinker when the weights are released, which the company expects next week if external testing finishes on schedule.