# Inkling 2 Small is capable, but its open weights have not shipped yet

Thinking Machines released a 594B-parameter mixture-of-experts model through Tinker and plans Apache 2.0 weights after safety testing.[1][2]

Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. The model has 594 billion total parameters, activates 17 billion for each token, accepts images and offers variable reasoning effort.[1][2]

The company made inference and fine-tuning available through Tinker at launch, while keeping the weights closed during external safety testing. It says it intends to publish the weights under Apache 2.0 next week. Treat that timing and license as a plan until the files and final terms are public.[1][2]

Independent measurements from Artificial Analysis support a large improvement over the first Inkling Small, but they also show long task times and high token use. Teams should compare accepted outcomes, elapsed time, review work and total cost on their own tasks before adopting it.[3][4]

## What Thinking Machines released

[Thinking Machines announced Inkling 2 Small on November 5 [1]](https://thinkingmachines.ai/news/inkling-2-small/). It is a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. It accepts text and images, returns text, supports variable effort and has a context window of up to one million tokens. Thinking Machines ran its launch evaluations with trajectories capped at the same length.[1][2]

People can sample and fine-tune the model through Tinker. The company also provides a Playground for trying it without setting up a training run.[1] Checkpoint downloads are tied to the planned weight release, so a Tinker account does not yet provide the complete model files.[1]

The launch price is $0.30 per million input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens.[2][3] Those rates cover hosted model tokens. Fine-tuning, long agent runs, storage and human review can change the cost of a completed task.

## Open weights are a promise until the files ship

Thinking Machines says it intends to release the full weights under Apache 2.0 next week after external safety tests and red-team work finish.[1][2] The model card describes the launch version as closed weights and gives no immutable download, checksum or final release commit.[2]

That distinction matters for procurement and research. Hosted access lets a team evaluate behavior and fine-tune through Tinker, but it does not yet permit local inference, offline inspection or independent redistribution of the promised files.

Wait for the actual repository, weight files, tokenizer, configuration, license text and checksums before recording the model as open weight. Recheck whether downloaded checkpoints include every component required for inference and whether Tinker-created checkpoints carry the same license.

## What the benchmark results establish

Artificial Analysis gives Inkling 2 Small a score of 31 on its Intelligence Index.[3][4] Its measurements show an increase from 5% to 46% on AutomationBench-AA and from 5% to 25% on FrontierCode 2.0 compared with Inkling Small.[3] These results support a material generation-to-generation improvement under that evaluator's harness.

The same evaluation shows mixed coding strengths. Inkling 2 Small scored 25% on FrontierCode 2.0, ahead of Claude Sonnet 5.5 at medium effort in that test, but scored 13% on Terminal-Bench 4.0 compared with Sonnet's 30%.[3] A single composite score can hide this task dependence.

Thinking Machines reports that Inkling 2 Small solved a mean 54.25 of 69 long-horizon coding tasks in an internal evaluation, while Claude Opus 5.5 solved 60.[1] The company also reports a four-and-a-half-hour autonomous research run that spent $426 on Tinker and more than doubled a forecasting score.[1] These are useful demonstrations, but outsiders still need the tasks, harness and repeated runs to assess generality.

## Latency and token use can outweigh the low token price

Artificial Analysis measured 116 output tokens per second and 34.54 seconds to the first answer token, but its full Intelligence Index tasks took 21.5 minutes on average.[4] As measured on November 6, comparable models in its selected group had a 7.3-minute median.[4]

The evaluator recorded about 150,000 output tokens per task, roughly four times Inkling Small's total.[3] The resulting cost was $1.18 per Intelligence Index task even though the published per-token rate is modest.[3][4]

Long output can be useful when an agent writes code, runs tools and checks its work. It can also add waiting time, review burden and more opportunities for an incorrect branch. Set explicit completion checks and token budgets instead of assuming that a lower active-parameter count makes every workflow cheap.

## A practical evaluation plan

- Choose 20 to 50 representative tasks that include ordinary cases, difficult cases and failures from your current system.
- Run Inkling 2 Small and the current model with the same tools, permissions, context, effort policy and stopping rules.
- Grade the finished artifact with executable checks where possible, then have a reviewer inspect correctness, security and fit for the user's request.
- Record accepted-task rate, elapsed time, input and output tokens, Tinker charges, tool failures, retries and reviewer minutes.
- Test image input separately with screenshots, diagrams and low-quality scans that resemble production data.
- For coding tasks, include repository navigation, test repair, dependency errors and requests that should stop for authorization.
- Repeat variable runs. A single successful demo cannot estimate reliability.
- If you plan local deployment, repeat performance and safety tests after the weights ship on the exact inference stack and quantization you intend to use.
- Keep the existing model as a fallback until the candidate passes quality, latency, cost and rollback thresholds.

## Limitations and open questions

- The company has announced an Apache 2.0 weight release, but the files and final release terms were not public when this guide was written.[1][2]
- Thinking Machines produced many of the launch evaluations and demonstrations.[1][2] Artificial Analysis provides independent measurements, but one benchmark suite cannot represent every workflow.[3][4]
- The one-million-token trajectory limit used in evaluations does not prove accurate recall across a one-million-token prompt.[1][2]
- The company reports stronger safety results than several open models, but local deployment will transfer monitoring, access control and incident response duties to operators.[2]
- Hardware requirements for useful local throughput depend on precision, expert routing, cache size, concurrency and the final released inference code.

## Frequently asked questions

- Can I download Inkling 2 Small now? No. Hosted sampling and fine-tuning are available through Tinker, while the full weight release is planned for next week.[1][2]
- Is it a 17B model? It activates 17 billion parameters per token, but the full mixture-of-experts model contains 594 billion parameters.[1][2]
- Does it support images? Yes. The model accepts image input through a 292-million-parameter vision encoder and returns text.[1][2]
- Is the model better than Claude for coding? The available results are mixed. It leads on some tests and trails on others, so evaluate the tasks and harness you plan to use.[1][3]
- What should I test first? Start with accepted outcomes, latency, total token use and reviewer effort on representative tasks. Repeat the tests after any local weight or inference-stack change.

## Sources

- [[1] Thinking Machines Lab: Introducing Inkling 2 Small](https://thinkingmachines.ai/news/inkling-2-small/)
- [[2] Thinking Machines Lab: Inkling 2 Small model card](https://thinkingmachines.ai/model-card/inkling-2-small/)
- [[3] Artificial Analysis: Inkling 2 Small release analysis](https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it)
- [[4] Artificial Analysis: Inkling 2 Small model measurements](https://artificialanalysis.ai/models/inkling-2-small)