Inkling 2 Small: 594B MoE, 17B Active, Live on Tinker (2026)

Inkling 2 Small is a Mixture-of-Experts (MoE) model from Thinking Machines Lab, announced on November 5, 2026, with 594B total parameters and 17B active per token. It accepts image input and does not support audio. You can use it today on Tinker, the company's training and inference platform, where the Playground handles inference and fine-tuning. The weights are closed for now, with a release under the Apache 2.0 license promised for "next week", and other platforms are expected "in the coming weeks". In short: for trying it out, Tinker is enough, and self-hosting can wait until the weights are published. This article covers the basic specs, what it can do, how it differs from the original Inkling, how to use Tinker and what it costs, vendor-reported benchmarks, a memory estimate for after the weights are released, a comparison with other large open-weight MoE models, and caveats, centered on facts from the official page.

Note that the official page gives no single maximum context length. It says pretraining used 512K and 1M contexts in later phases, reinforcement learning used 256K with compaction for longer trajectories, and benchmarks allowed trajectories of up to 1M tokens. Tinker's pricing documentation, meanwhile, lists a 64K standard version and a 256K version on separate rows. When using it, rely on the context length shown in the interface you use.

The official page emphasizes three areas. The first is visual understanding including image input, with examples such as reading charts and screens and building websites from images. The second is coding, from fixing repositories to research-style work such as GPU kernel optimization. The third is office and knowledge-work agents, with results shown on document, tool-use, finance and legal benchmarks.

- Image input: reading screens and charts, identifying things from photos, building web apps from visual cues
- Coding: fixing real repositories (SWE-style), terminal work, automating research tasks
- Agent work: tool calling, work through MCP, document, spreadsheet and finance tasks
- Long tasks: reinforcement learning at 256K, with compaction for longer trajectories

The page is also candid about weaknesses: it is not the most succinct communicator, it may not pick the best tool in codebases with many skills, it has no audio support, it is less optimized for chat, and it trails top closed-weight models on safety and alignment evaluations.

The original Inkling was Thinking Machines Lab's first open-weight model, announced in July 2026. It is an MoE with 975B total and about 41B active parameters, accepts text, images and audio, supports a context of up to one million tokens, and uses the Apache 2.0 license. For its memory requirements, see our Inkling requirements guide.

By the numbers, total parameters are about 60% of the original and active parameters a little over 40%. Even so, the official page says it far surpasses the original on the Artificial Analysis Intelligence Index. Audio, however, was dropped. Inkling 2 Small can therefore be read as a generation that is cheaper and faster for uses that do not need audio. The official page does not state the original's parameter count, so the original's values in the table come from our existing article.

On usability, the official page says instruction following, careful weighing of changes and unprompted verification of its own work have improved. Alignment metrics also improved: for example, an internal misbehavior metric fell from 66.8 for the original to 14.9, and a destructive-actions metric fell from 71.9 to 4.7. All of these are the vendor's own evaluations.

Tinker is currently the only way in. From "Try on Tinker" on the official page you open the Tinker Playground and select Inkling 2 Small as the model. The Playground lets you try inference (sampling) and fine-tuning, and checkpoints are said to become downloadable when the weights are released.

- Step 1: Prepare a Tinker account and open the Playground
- Step 2: Select Inkling 2 Small from the model list and check its behavior with prompts
- Step 3: For images, attach image input (audio is not available)
- Step 4: To tune it on your own data, use Tinker's fine-tuning
- Step 5: After the weights are released, begin considering fetching the checkpoint and moving it to your own environment

For programmatic use, the model identifiers in Tinker's pricing documentation are thinkingmachines/Inkling-2-Small (64K context) and thinkingmachines/Inkling-2-Small:peft:262144 (256K variant). The Tinker overview page we could check had no setup steps or code examples, so this article does not show concrete SDK calls. Please check the official documentation (tinker-docs.thinkingmachines.ai) for current instructions.

The official announcement itself lists no prices and points to Tinker's models and pricing documentation. There, Inkling 2 Small is listed as follows (all per million tokens, as checked on November 7, 2026).

Only training costs more for the 256K variant; prefill and sampling are the same. The same table lists the original Inkling's sampling at $4.68 and Inkling-Small at $1.44, so Inkling 2 Small is near the lowest sampling price there. Separately, serverless inference (beta) is listed as thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4 at $1.00 output and $0.30 input (cached $0.03), but it is noted as waitlist-only. Prices may change, so check the latest page before use.

Everything below is a value published by Thinking Machines Lab on its official page, not an independent reproduction. According to the page, benchmarks were run at effort 0.99 (xhigh) and temperature 1.0.

A few reading notes. First, these are self-published values, and evaluation settings such as effort and maximum trajectory length affect results. Second, the page itself acknowledges that it trails GPT-6.1 Sol on alignment evaluations. Third, some absolute scores are low (FrontierCode at 25.4%, HLE-Diamond at 28.7%), so it is not strong across the board. Whether it works for your use case has to be judged on your own tasks.

The weight release is still only announced, and the distribution format (BF16, NVFP4 or a quantized build) could not be confirmed at the time of writing. We therefore give a rough estimate as a simple calculation from the 594B total parameter count. These are not official figures.

In real operation, KV cache, activations, runtime buffers and the vision encoder (292M) come on top. The longer the context, the larger the KV cache, so adding roughly 20% headroom to the table values is realistic. One more point is easy to confuse: what 17B active means. With 17B active, per-token compute is small and generation tends to be fast. But in an MoE you cannot know which experts will be selected, so the entire set of weights must stay in memory. Speed follows active parameters; required memory follows total parameters. At 594B, this is not the scale of a personal PC or an ordinary workstation, but of a multi-GPU server or a machine with large unified memory. You can also use our VRAM calculator for estimates, but since this model is not yet released, it is not registered there.

By total parameters, Inkling 2 Small is considerably smaller than the other large open-weight MoE models. The comparison below includes only figures confirmed in our existing articles or on official pages; unconfirmed items are shown as a dash.

With fewer total parameters, the weights are likely to be smaller, so after release it may become a large model with a relatively lower bar for self-hosting. The actual size, however, depends on the distribution format. Also, Tinker's sampling price and other vendors' API prices have different billing terms, so they should not be lined up too directly.

- The weights are not yet released. The Apache 2.0 release is a promise, and timing or contents could change depending on external safety testing
- All benchmarks are vendor self-reports. Wait for independent reproductions before judging
- Audio is not supported. Uses that relied on the original Inkling's audio input cannot migrate
- The official page admits chat optimization is weaker. Test first if conversational UI quality matters
- The memory figures in this article are simple calculations, not official values, and will be replaced by measured values after the weights are released
- The official page also describes a case in which, during reinforcement learning, the model tried to work around a network block through a third-party service. If you connect it to external tools as an agent, you need to design permissions and monitoring