Thinking Machines publishes Inkling 2 Small weights after hosted-only debut

Thinking Machines Lab made Inkling 2 Small’s weights available to download on November 12, 2026, ending the model’s initial hosted-only phase. The release puts the trained model files on Hugging Face, alongside new serving and training integrations and a wider set of API providers. Together AI lists the model under Apache 2.0.

The download follows a week of closed weights

The lab introduced Inkling 2 Small on November 5 through Tinker, where developers could run and fine-tune it without downloading the weights. At launch, the lab said it intended to publish the full weights the following week after external safety testing and red-teaming—adversarial attempts to expose problems. The November 12 update confirms that downloads are available; it does not describe the results of that testing.

There are two dates to keep separate. Together’s catalog gives November 4 as the model’s release date, while the lab’s introduction is dated November 5. Neither is the date of the newly confirmed weight download: that change appears in the lab’s November 12 update.

“Small” describes only part of the model

Inkling 2 Small contains 594 billion parameters—the learned values inside the model—but activates 17 billion for each token it processes. Its mixture-of-experts design routes each token through eight of 384 specialized components. The active count describes the portion used for a token, not the size of the complete model.

It accepts text and images and has a one-million-token context window, the amount of material it can work with at once. Reasoning effort is adjustable per request. Together describes that control as a way to balance speed against thoroughness, rather than using the same effort setting for every task.

For developers choosing hosted access, Together offers serverless and dedicated infrastructure at $0.30 per million input tokens and $1 per million output tokens. Its API identifies the model as thinkingmachines/Inkling-2-Small. Those are Together’s rates, not a price for downloading or operating the checkpoint elsewhere.

The lab’s update also lists API access through NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare. On the software side, it names integrations with Unsloth, SGLang, vLLM and TokenSpeed, plus post-training support through Miles, SkyRL and Snowflake Cortex Training. The announcement therefore covers both access to the files and tools for working with them.

Wider access does not mean a clean benchmark sweep

Thinking Machines reports an Artificial Analysis Intelligence Index score of 31, substantially above its previous models. Its coding comparisons are less uniform. In the lab’s published table, Inkling scores 72.1% on DeepSWE v1.1, behind DeepSeek-V4.1-Flash-1017 at 77%. On FrontierCode v2.0, the corresponding scores are 25.4% and 28.5%.

The same table puts Inkling ahead of GLM-5.3-Flash, Muse Spark Open and Kimi K3 on both tests. That is a competitive position, not an across-the-board lead. These are company-published comparisons; Together says the reported results use reasoning effort of 0.99, labeled xhigh, and temperature 1.0.

The capability pitch extends beyond code. Thinking Machines says the model can use visual feedback to inspect and refine websites, games and 3D designs. Its coding training included hundreds of thousands of long-running tasks, with some graders assessing how the model approached a problem rather than only whether the resulting code passed tests.

Together’s model card describes a unified reinforcement-learning run spanning coding, vision, knowledge work, tool use and safety. It also lists training environments that rewarded stopping on impossible tasks. Those are elements of the model’s training design, distinct from the external safety testing that Thinking Machines originally placed before the downloadable release.