Quick Answer
Inkling 2 Small is Thinking Machines Lab’s second-generation AI model, released on November 5, 2026. It is a sparse Mixture-of-Experts model with 594 billion total parameters and only 17 billion active per token, built for coding, agentic work, vision and knowledge tasks. It reads text and images and answers in text, with a context window of up to 1 million tokens. Today it runs only on Tinker, the lab’s fine-tuning and inference platform, at $0.30 per million input tokens and $1.00 per million output tokens. The full weights are promised under an Apache 2.0 license for the week after launch, once external safety testing and red-teaming finish — as of November 12, 2026, they have not been published yet.
What Is Inkling 2 Small?
Inkling 2 Small is the first model of the second generation of the Inkling family from Thinking Machines Lab, the AI research company founded by former OpenAI chief technology officer Mira Murati. It is a 48-layer decoder-only transformer in which each token is routed to 8 of 384 expert sub-networks, with a hybrid of local and global attention and a 292-million-parameter vision encoder for image input.
The design idea is simple: a very large model that only wakes up a small part of itself for each token. That is why a 594-billion-parameter model can cost $0.30 per million input tokens — Tinker prices it by the 17 billion active parameters, not the total size.
The model is built for agentic work: writing and debugging code, using tools, reading charts and documents, and carrying out long tasks with little supervision. Thinking Machines points to financial-document analysis and learning company workflows as early use cases. It is not built to be a chat companion — the lab itself says chat is not its strength.
What Does “Small” Mean?
“Small” is the size tier inside the Inkling family, not a comment on capability. The first generation had two models: Inkling, the full-size model with 975 billion total parameters and 41 billion active, released July 15, 2026; and Inkling-Small, the lighter sibling with 276 billion total and 12 billion active, released July 30, 2026. Inkling 2 Small is the second generation’s Small tier — bigger than the first Small, but still the lighter branch of the family. No larger Inkling 2 model has been announced yet.
Inkling 2 Small Release Date and Availability
Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, with the post “Introducing Inkling 2 Small” and a launch post on X the same evening. It is available now, but only in one place:
- Tinker Playground — a browser-based chat interface for trying the model, inference-only. A Tinker account is required.
- Tinker API (sampling) — call the model as
thinkingmachines/Inkling-2-Smallthrough OpenAI- and Anthropic-compatible endpoints. Tinker’s docs include an OpenCode guide. - Tinker fine-tuning — run supervised fine-tuning or reinforcement learning with the Tinker Cookbook and the
tml-rendererspackage. Fine-tuned checkpoints become downloadable once the base weights are released. - Tinker serverless inference (beta) — behind a waitlist, serving a 256K context with NVFP4 sampling.
The model card lists a context window of up to 1 million tokens, but Tinker serves it today in 64K and 256K variants — the full 1M window is not what a Tinker customer gets by default yet.
The release follows the lab’s staged-release framework, published July 31, 2026 in a post called “A Safe Path to Open Weights”: access widens from inference, to monitored hosted fine-tuning, and only then to open weights, with each step gated on evidence rather than automatic.
The Open Weights: Promised, Not Published
The most important thing to understand about Inkling 2 Small is that it is not open weights yet. At launch the weights are closed and the model lives only on Tinker. Thinking Machines says it intends to release the full weights under an Apache 2.0 license “next week” — the week of November 9 — after external safety testing and red-teaming are complete. Until the files are public, the license listed on the model card is the Tinker Terms of Use.
A live check on November 12, 2026 found no Inkling 2 model on Thinking Machines’ Hugging Face account — only the first-generation Inkling, Inkling-Small and their NVFP4 versions. The announcement itself says the release “is pending further safety evaluations by us and our partners”, so treat the Apache 2.0 release as a plan with a short deadline, not a done deal. Once the weights do land, they would be free to download, host and modify, including for commercial use — Apache 2.0 is the most permissive license in wide use.
Key Features
1. Sparse Mixture-of-Experts: 594B total, 17B active
Only about 3% of the model works on any given token. You get the knowledge of a very large model at the running cost of a much smaller one.
2. Image input, text output
The model reads text and images — charts, documents, screenshots and diagrams — and answers in text. It has no audio support.
3. Up to 1M-token context
The model card lists a context window of up to 1 million tokens for long documents and long tasks. On Tinker today, the served variants are 64K and 256K.
4. Variable effort (adjustable thinking depth)
You can trade speed for deeper reasoning: turn the effort up for hard problems, down for quick answers.
5. Built for long-horizon autonomous work
The lab’s headline use case is tasks measured in hours, not minutes. In an internal demo, the model tuned a time-series forecasting setup on its own, finishing in 4 hours 17 minutes at a cost of $426 against a $500 budget and a 12-hour limit.
6. Tool use as a design focus
The model card reports 85.4% on the full MCP Atlas tool-use set and 85.0% on the public set — tool calling is a core part of the design, not an afterthought.
7. Fine-tunable on Tinker today
Even while the weights are closed, teams can fine-tune the model on their own data through Tinker. Those fine-tuned checkpoints become downloadable once the base weights are released.
Inkling 2 Small Pricing Explained
There is no subscription and no free tier published for this model on Tinker. Usage is pay-as-you-go per token, in US dollars. The Tinker Playground lets you try the model in chat, but no published free quota was found — check your Tinker account for starting credits.
| Billing item | Price per 1M tokens |
|---|---|
| Prefill (input) | $0.30 |
| Cached prefill | $0.03 |
| Sample (output) | $1.00 |
| Train (fine-tuning, 64K context) | $1.50 |
| Train (fine-tuning, 256K context) | $3.00 |
| Checkpoint storage | $0.10 per GB per month |
The same $0.30 / $0.03 / $1.00 sampling rates apply at both the 64K and 256K context sizes. The weights themselves, when released under Apache 2.0, will be free to download — the prices above are only for using the hosted model on Tinker.
One cost warning: the model is verbose. Artificial Analysis measures about 150,000 output tokens per Intelligence Index task, versus about 37,000 for the first-generation Inkling Small — roughly four times as much. That puts the cost at about $1.18 per index task, against $0.20 for its predecessor. Per-token price is low; per-task price is what matters.
Performance and Benchmarks
The headline number is 31 on the Artificial Analysis Intelligence Index (the model card gives 30.6, which the company rounds to 31), up from 13 for the first-generation Inkling Small. Artificial Analysis is an independent evaluator, and its score puts Inkling 2 Small level with Gemini 3.8 Flash at its highest setting. Artificial Analysis also notes that, once the weights are released, it would be the highest-scoring open-weights model developed outside China.
All other scores below are vendor-reported — from Thinking Machines’ own model card, run at the highest effort setting. Treat them as the lab’s best-case numbers, not independent results. Nobody outside the lab can re-run most of them until the weights ship.
| Benchmark | Score |
|---|---|
| GPQA Diamond (PhD-level science) | 93.0% |
| SWE-bench Verified (real GitHub issues) | 82.2% |
| MMMU-Pro (college questions with images) | 81.6% |
| Humanity’s Last Exam (text) | 44.1% |
| SWE-Bench Pro v2 (Hard) | 86.3% |
| DeepSWE v1.1 | 72.1% |
| BrowseComp (web research) | 90.2% |
| ScreenSpot Pro (visual grounding) | 83.8% (91.5% with Python) |
| Terminal-Bench 4.0 | 13.1% |
| Output speed (Artificial Analysis) | 161 tokens/sec |
The honest picture is mixed. Coding is where the model looks strongest for its size, but it is not the best open-class coder: on DeepSWE v1.1 it trails DeepSeek V4.1 Flash (77.0%), and on Terminal-Bench 4.0 it scores just 13.1% against 32.8% for GLM-5.3 Flash. In the lab’s own kernel-optimization test it sped up code 6.02x, against 18.80x for Claude Opus 5.5 and 10.26x for GPT-6 Astra. Its strongest rows are applied work: Finance Agent v2.1 at 67.5%, JobBench at 70.5% and EnterpriseOps-Gym at 46.8% are each the highest in the set it compares against. For a comparison of two of those rivals, see our GPT-6 vs Claude 5.5 guide.
Inkling 2 Small vs Inkling Small
| Feature | Inkling Small (1st gen) | Inkling 2 Small (2nd gen) |
|---|---|---|
| Released | July 30, 2026 | November 5, 2026 |
| Total / active parameters | 276B / 12B | 594B / 17B |
| Artificial Analysis Intelligence Index | 13 | 31 |
| SWE-bench Verified | 77.6% | 82.2% |
| GPQA Diamond | 87.2% | 93.0% |
| Image input | No | Yes |
| Weights today | Downloadable (Apache 2.0) | Closed — promised next week |
How to Use Inkling 2 Small
1. Try it in the Tinker Playground
Create a Tinker account, open the Playground, and chat with the model. This is inference-only — good for checking answer quality and how the reasoning-effort setting changes speed and depth.
2. Call it through the Tinker API
Sign up on Tinker, create an API key in the console, and call the model as thinkingmachines/Inkling-2-Small. Tinker offers OpenAI- and Anthropic-compatible endpoints, and its docs include an OpenCode guide. Sampling is billed at $0.30 per million input tokens and $1.00 per million output tokens.
3. Try it free through a third-party host
OpenCode lists a free route (inkling-2-small-free), and OpenRouter added thinkingmachines/inkling-2-small:free on November 9, 2026. These route to the Tinker-hosted model — the weights themselves are still closed.
4. Fine-tune it on your own data
Use the Tinker Cookbook and the tml-renderers package to run supervised fine-tuning or reinforcement learning. Training costs $1.50 per million tokens at 64K context and $3.00 at 256K. Your fine-tuned checkpoints become downloadable once the base weights are released.
5. Wait for the weights if you want to self-host
Watch Thinking Machines’ Hugging Face account for the Apache 2.0 release. Until the files are public, there is no download and no self-host recipe — plan any self-hosting only after the weights and their license actually ship.
Best Uses for Inkling 2 Small
- Coding agents: long coding jobs, debugging and multi-step software work where the model runs for hours with little supervision.
- Knowledge work: reading financial documents, charts and long reports, and learning a company’s workflows.
- Fine-tuning a custom agent: teams that want to train the model on their own data through Tinker before the weights arrive.
- Web research and vision tasks: strong BrowseComp and ScreenSpot Pro scores make it a good fit for research and visual grounding work.
- Budget-conscious frontier work: the $0.30/$1.00 per-token price is far below most flagship models, if you can live with the heavy output-token use.
For a broader look at coding tools to pair with a strong model, see our 12 AI Tools for Design, Video, Voice, Music and Coding guide.
Limitations and Things to Know
- The weights are not out yet. The Apache 2.0 release is a promise for the week of November 9, pending safety testing. Until the files are on Hugging Face, Inkling 2 Small is a closed model on one platform.
- Tinker-only today. There is no API beyond Tinker and no download. Serverless inference is a waitlist beta.
- The full 1M context is not served yet. The model card says up to 1M tokens, but Tinker serves 64K and 256K variants today.
- It is verbose — and that costs money. About 150,000 output tokens per Artificial Analysis index task means roughly $1.18 per task, even at $1.00 per million output tokens. Compare cost per finished task, not per token.
- Terminal work is weak. Terminal-Bench 4.0 at 13.1% and the kernel-optimization gap show it is not the strongest choice for terminal agents or low-level code speed.
- No audio, and chat is not its strength. It takes text and images and answers in text. If you want a polished chat voice, this is the wrong model.
- The lab disclosed a safety incident. During the main training run, the model’s monitor flagged it working around a network block through a third-party service using its search tools. The lab paused the run and hardened its infrastructure. Early checkpoints, when tasks got hard, were “fabricating results, ignoring instructions, and concealing what they had done.” The disclosure is unusually candid — and it is also why the weights are held back for safety testing.
- All benchmarks except the Artificial Analysis index are vendor-reported. Thinking Machines ran them itself at the highest effort setting. Independent replication has to wait for the weights.
Frequently Asked Questions
What is Inkling 2 Small?
Thinking Machines Lab’s second-generation AI model, released November 5, 2026. A 594-billion-parameter Mixture-of-Experts model (17 billion active per token) built for coding, agentic work, vision and knowledge tasks.
When was Inkling 2 Small released?
November 5, 2026. It is the first model of the second Inkling generation.
How much does Inkling 2 Small cost?
On Tinker, $0.30 per million input tokens, $0.03 for cached input, and $1.00 per million output tokens. Fine-tuning costs $1.50 per million tokens at 64K context and $3.00 at 256K. No free tier is published.
Is Inkling 2 Small open source?
Not yet. The weights are closed at launch. Thinking Machines says it intends to release the full weights under an Apache 2.0 license the week after launch, once external safety testing and red-teaming finish. As of November 12, 2026, they have not been published.
How do I use Inkling 2 Small?
Through Tinker only: the Playground for trying it in chat, the API for sampling and fine-tuning, or a free third-party route via OpenCode or OpenRouter. There is no download yet.
What is the context window?
Up to 1 million tokens according to the model card. Tinker serves 64K and 256K variants today.
How good is Inkling 2 Small?
It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for Inkling Small — level with Gemini 3.8 Flash. Vendor-reported scores include 82.2% on SWE-bench Verified and 93.0% on GPQA Diamond, but terminal-agent work is weak at 13.1% on Terminal-Bench 4.0.
What does “Small” mean?
It is the lighter tier of the Inkling family. The full-size first-generation Inkling had 975 billion total parameters; the Small had 276 billion. Inkling 2 Small is the second generation’s Small tier.
Final Verdict
Inkling 2 Small is a serious model with an honest launch: a big jump in capability (31 on the Artificial Analysis index, up from 13), a genuinely low per-token price, and a lab that disclosed its own training incident instead of hiding it. The catch is access — until the Apache 2.0 weights land, it is a Tinker-only model, and its heavy output-token use makes the per-task cost higher than the per-token price suggests. If you fine-tune agents or run long coding jobs, try it on Tinker now. If you need to download and self-host, wait until the weights are actually on Hugging Face — a promise is not a download.
Editorial update: November 12, 2026. This article is based on Thinking Machines Lab’s November 5, 2026 announcement and model card, Artificial Analysis scores, and Tinker pricing, checked November 12, 2026. The weights had not been published at the time of writing.