Thinking Machines Lab released Inkling 2 Small on November 5, 2026, and it is the lab’s second model generation. Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters and only 17 billion active for each token, so it stays cheap to run while it tackles long coding jobs. This guide covers what the model does well, where it falls short, what it costs on Tinker, and how to try it today.
The weights are still closed at launch, and the lab plans an Apache 2.0 release next week. Therefore, the next few days are the right moment to test the hosted version.
What Is Inkling 2 Small?
Inkling 2 Small is the first model of Thinking Machines Lab’s second generation. The lab was founded by former OpenAI chief technology officer Mira Murati, and it builds open-weight models that teams can customize with their own data. Its first generation arrived in July 2026 with Inkling, a 975 billion parameter model. The new model replaces that first generation, and it targets coding, autonomous software work, and developer customization rather than simple chat.
The model accepts text and image inputs, and it supports variable reasoning effort, so you can ask it to think harder on difficult tasks and answer fast on simple ones. In addition, it handles up to 1 million tokens of context on its model card, although Tinker serves it today in 64,000 and 256,000 token variants. Training covered 8.4 trillion tokens in the first phase, followed by reinforcement learning on long-horizon coding tasks.
The design is a sparse mixture of experts: a router sends each token to 8 of 384 experts, so only 17 billion parameters work on any single token. For example, a business pays for the active compute, not for the full model size. However, the design is only as good as the router, and the lab says the second generation improves routing enough to lift scores.
Inkling 2 Small Pricing on Tinker
Inkling 2 Small is available today only through Tinker, the lab’s platform for inference and fine-tuning. The serverless beta charges $0.30 per million input tokens, $0.03 per million cached input tokens, and $1.00 per million output tokens. Training costs $1.50 per million tokens at a 64,000 token context and $3.00 at a 256,000 token context. Production use sits behind a waitlist.
| Tinker service | Price |
|---|---|
| Input tokens | $0.30 per million |
| Cached input tokens | $0.03 per million |
| Output tokens | $1.00 per million |
| Training, 64K context | $1.50 per million |
The per-token price looks cheap, and in isolation it is. Output costs about 79 percent less than the original Inkling model. However, the real cost depends on how much the model writes. Artificial Analysis measured about 150,000 output tokens per index task, roughly four times the first Inkling Small, so a single task cost $1.18 even at these low rates. In contrast, Mistral Large 4 cost $2.06 per task, so the new model still wins per task. Meanwhile, teams should budget for verbosity, because long answers can erase a per-token saving fast.
Inkling 2 Small Benchmarks: The Wins and the Weakness
Benchmarks tell a mixed but encouraging story. On the Artificial Analysis Intelligence Index version 5.0, the model scores 31, a large jump from 13 for the first Inkling Small, and it ranks fourth of 29 models in its price class. The same evaluator puts it level with Google DeepMind’s Gemini 3.8 Flash. In addition, independent checks confirm the jump is real even where the model does not lead.
| Benchmark | Score | Closest rival |
|---|---|---|
| Artificial Analysis Index v5.0 | 30.6 | DeepSeek V4.1 Flash 1017, 32.6 |
| SWE-Bench Pro v2, Hard | 86.3% | Kimi K3, 88.2% |
| DeepSWE v1.1 | 72.1% | Mistral Large 4, 61.7% |
| Vals Finance Index | 61.5% | DeepSeek V4.1 Flash 1017, 60.9% |
| Terminal-Bench 4.0 | 13.1% | DeepSeek V4.1 Flash 1017, 33.8% |
Where Inkling 2 Small Leads
The coding results are the strongest part of the story. The model scores 86.3 percent on SWE-Bench Pro version 2 Hard, second only to Kimi K3 at 88.2 percent, and it reaches 72.1 percent on DeepSWE version 1.1, well ahead of Mistral Large 4 at 61.7 percent. It also jumps from 5 percent to 46 percent on AutomationBench-AA and from 5 percent to 25 percent on FrontierCode 2.0.
Where It Lags
Terminal-Bench 4.0 is the clear weak spot, because the model scores only 13.1 percent there while DeepSeek V4.1 Flash 1017 reaches 33.8 percent. Speed is the second concern: Artificial Analysis measured 116 output tokens per second and 34.54 seconds to the first answer token, and its full index tasks took 21.5 minutes on average against a 7.3 minute median for peers. Therefore, treat this model as a deep worker for long jobs, not as a fast responder for live chat. The lab chose the comparisons in its own table, so treat the wins as company-framed until independent tests repeat them.
Inkling 2 Small vs DeepSeek V4.1 Flash
DeepSeek V4.1 Flash 1017 is the natural rival, because both models chase cheap, capable coding work and both publish open weights. DeepSeek leads on the Artificial Analysis Index at 32.6 against 30.6, and it dominates Terminal-Bench at 33.8 percent against 13.1 percent. In addition, DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens, so it undercuts the new model on raw price. On the other hand, Inkling 2 Small leads on vision, finance, and hard reasoning in the lab’s own tables, with 61.5 percent on the Vals Finance Index against DeepSeek at 60.9 percent.
The choice depends on your workload. If your agents live in the terminal and run shell commands all day, DeepSeek is the safer pick today, and our DeepSeek Super Agent guide explains that ecosystem in detail. If your work is long-horizon coding, web building, or finance analysis, and you want a United States built open-weight model with first-party fine-tuning, the new model is the stronger bet.
How to Try It on Tinker
Getting started takes minutes, because the Tinker Playground lets you try the model without setting up a training run. You need a Tinker account, and then you can sample the model and compare it with your current stack. The Inkling 2 Small Tinker API also supports sampling and fine-tuning. Follow these steps to run a fair first test.
- Create a Tinker account and open the Playground.
- Run the same coding task on both models, then compare the answers.
- Measure time to first token, output tokens, and cost per task.
- Try a long-horizon task, such as building a small web app.
- Fine-tune on your own codebase only after the Playground results justify the spend.
Open Weights: What the Apache 2.0 Plan Means
The weights are closed at launch, and that fact shapes every decision this week. Thinking Machines Lab says it will release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming finish. Treat that timing as a plan, not a promise: the model card still lists the launch version as closed weights. If the release lands as planned, Artificial Analysis says it would be the highest-scoring open-weight model developed outside China, and named runtimes include Ollama, vLLM, LM Studio, llama.cpp, Unsloth, and SGLang.
An Apache 2.0 release changes the economics completely. Teams could self-host the model, fine-tune it without platform lock-in, and run it on private infrastructure for sensitive data. That is the same promise in our NVIDIA Nemotron 3.5 guide, where local deployment keeps client data inside the office. In addition, the first Inkling generation already ships open weights, so the lab has a track record of following through. Meanwhile, test on Tinker now and hold the self-hosting budget until the files and final terms are public.
Inkling 2 Small for Business Agents
Business agents are the real audience for this model, because it is built for long, multi-step work rather than quick answers. Thinking Machines Lab reports a four-and-a-half-hour autonomous research run that spent $426 on Tinker and more than doubled a forecasting score. That single run shows both the promise and the cost: the model can sustain a long task, but the bill grows with every token it writes. As a result, the best business uses are jobs where a long, careful run is worth more than a fast, cheap one: coding agents, support routing, and document analysis.
For example, a team could let the model draft a full feature, run the tests, and fix the failures in one session. Our Mercury 3 guide covers a diffusion model that trades in the opposite direction, with far faster output for real-time agents. In contrast, this model trades speed for depth, so pair it with a fast model for live replies and reserve it for the heavy work behind the scenes.
Limitations You Should Know
No model is right for every job, and this one has clear limits. Firstly, the weights are closed today, so you cannot self-host until the Apache 2.0 release. Secondly, Terminal-Bench performance is weak, so terminal-heavy automation is a poor fit. Thirdly, the model is slow to start and verbose, with 34.54 seconds to first token and about 150,000 output tokens per index task. Finally, every headline benchmark comes from the lab’s own tables, so independent replication still matters. Test your exact workload, measure cost per finished task, and keep your current model as a fallback.
Frequently Asked Questions
What does it cost on Tinker?
Inkling 2 Small pricing is $0.30 per million input tokens, $0.03 per million cached input tokens, and $1.00 per million output tokens. Training costs $1.50 per million tokens at a 64,000 token context.
How good are the benchmarks?
Inkling 2 Small benchmarks look strong on coding and finance: 31 on the Artificial Analysis Intelligence Index, 86.3 percent on SWE-Bench Pro version 2 Hard, and 72.1 percent on DeepSWE version 1.1. However, Terminal-Bench 4.0 is weak at 13.1 percent.
Is it open weights yet?
No, the weights are closed at launch. The Inkling 2 Small open weights release is planned under Apache 2.0 in the week of November 10 to 14, after safety testing and red-teaming. Until the files are public, the model runs only on Tinker.
How do I use the Tinker API?
Create a Tinker account, then use the Playground for tests or the API for sampling and fine-tuning. Production access sits behind a waitlist.
Which is better for coding, this model or DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash leads on terminal coding and raw price, while Inkling 2 Small leads on long-horizon coding, vision, and finance. Choose DeepSeek for shell-heavy automation and the new model for sustained coding projects.
Can small businesses use it for agents?
Yes, small businesses can use Inkling 2 Small for business agents such as coding agents, support routing, and document analysis. Start with one pilot workload, measure cost per finished task, and keep a fast model for live replies, because this model is slow to start.
Who built this model?
Thinking Machines Lab built it. The lab was founded by former OpenAI chief technology officer Mira Murati, and this is its second model generation after the first Inkling models of July 2026. Learn more in the official Thinking Machines Lab announcement and the MarkTechPost launch report.
Conclusion
Inkling 2 Small is a serious second-generation model: a 594 billion parameter mixture-of-experts system that trades speed for depth on long coding and agent work. It scores 31 on the Artificial Analysis Intelligence Index, reaches 86.3 percent on SWE-Bench Pro Hard, and costs $0.30 per million input tokens on Tinker. The Terminal-Bench weakness and the slow first token are real limits, and the closed weights mean this week is a testing window, not a deployment deadline. Try it in the Playground today, measure cost per finished task, and watch for the Apache 2.0 release next week.
