AI Basics & TutorialsThinking MachinesInkling 2 SmallTinkeropen weightsMira Murati

Thinking Machines Inkling 2 Small Costs $0.30 Per Million Tokens: 594 Billion Parameters, and the Weights Land Next Week

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a 594 billion parameter mixture-of-experts model that costs $0.30 per million input tokens and $1.00 per million output tokens on the company's Tinker platform. The weights are closed today, with an Apache 2.0 release promised for next week after safety testing.

Published November 10, 2026

Muhammad Usman

By Muhammad Usman · Founder & Lead Reviewer

Thinking Machines Inkling 2 Small Costs $0.30 Per Million Tokens: 594 Billion Parameters, and the Weights Land Next Week

Quick Answer

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a 594 billion parameter mixture-of-experts model that costs $0.30 per million input tokens and $1.00 per million output tokens on the company's Tinker platform. The weights are closed today, with an Apache 2.0 release promised for next week after safety testing.

Key takeaways:

  • Inkling 2 Small launched November 5, 2026: 594 billion total parameters, 17 billion active per token, text and image input.
  • It costs $0.30 per million input tokens, $0.03 cached, and $1.00 per million output tokens on Tinker serverless inference.
  • It scores 31 on the Artificial Analysis Intelligence Index, the highest score the lab has reported for its models.
  • The weights are not public yet. Thinking Machines says they ship under Apache 2.0 next week, once external safety testing and red-teaming finish.
  • Today you can only reach it through Tinker, which serves 64K and 256K context, not the full 1 million tokens in the model card.

What is Inkling 2 Small?

Inkling 2 Small is the first model of the second generation from Thinking Machines Lab, the company founded by Mira Murati in February 2025. It launched on November 5, 2026, on Tinker, the company's fine-tuning and inference platform, which launched in November 2025. The first Inkling generation arrived on July 15, 2026, so the second generation follows less than four months later.

The model is a sparse mixture-of-experts transformer with 594 billion total parameters, of which 17 billion are active for each token. Each token is routed to 8 of 384 experts. It takes text and images as input and returns text, with variable inference effort, so you can ask it to think harder on difficult prompts and lighter on easy ones. The model card lists a context window of up to 1 million tokens, although what you actually get on Tinker today is smaller, which we cover below.

Thinking Machines is the lab to watch here for a simple reason: it is the most credible independent challenger to the big labs, staffed by former OpenAI and DeepMind researchers, and it is building in the open. Inkling 2 Small is its statement that a 594 billion parameter model can be served cheaply, with open weights to follow.

What does it cost, exactly?

Tinker's serverless inference, which is in beta, charges $0.30 per million input tokens, $0.03 per million cached input tokens, and $1.00 per million output tokens. That cached input rate is a 90 percent discount on the input rate, so repeated context costs almost nothing.

Fine-tuning is also priced on the page. Training costs $1.50 per million tokens at 64K context and $3.00 per million tokens at 256K context. Checkpoint storage costs $0.10 per GB per month. Tinker prices its mixture-of-experts models by active parameters rather than total size, which is why a 594 billion parameter model can cost $0.30 per million input tokens: you are paying for the 17 billion active parameters, not the full 594 billion.

To put that in perspective, a job of 1 million input tokens and 250,000 output tokens costs $0.30 plus $0.25, or $0.55. The same job on Claude Opus 5.5 at $4 input and $20 output would cost $4.00 plus $5.00, or $9.00. That is the gap that makes Inkling 2 Small interesting: frontier-class size at a price that lets you run real workloads without watching the meter.

How good is it?

The headline number is 31 on the Artificial Analysis Intelligence Index, the highest score the lab has reported for its models. One briefing describes it as the highest-scoring open-weights model outside China, once the weights ship. That claim depends on the weights actually shipping, so treat it as a promise, not a fact, until the release lands.

Artificial Analysis also publishes a cost-per-task figure for running its index: $1.18 per task for Inkling 2 Small, against $2.06 for Mistral Large 4 and $2.02 for Muse Spark. That is the cost of the benchmark work itself, not a general price for your workload, but it shows the same pattern as the per-token rates: Inkling 2 Small is cheap to run for the quality it reports.

Be careful with the rest of the benchmark story. Thinking Machines has not published a full independent benchmark suite for Inkling 2 Small in the material checked for this article, and the figures available are the lab's own or from index providers. There is no independent evaluation yet, and no published result on the hardest coding or reasoning suites that would let you compare it directly with the frontier closed models. The honest summary is that the price is verified, the index score is verified, and the rest is still to be proven.

Where can you use it today?

Only on Tinker. Inkling 2 Small is available today for hosted inference and fine-tuning through Tinker's playground and its OpenAI-compatible API, which supports chat completions and completions. There is no download, no weights file, and no other platform serving it yet. Thinking Machines says other platforms will follow in the coming weeks.

Two limits matter. First, Tinker serves the model in 64K and 256K context variants today, so the full 1 million token window in the model card is not what a developer gets by default on the platform. If your work needs a million tokens of context, wait for the weights or for Tinker to expose the longer variant. Second, Tinker's serverless inference is in beta, with a waitlist for production use, and no status page or service level agreement has been published. That is fine for experiments and small jobs. It is not a platform to bet a production deadline on yet.

When do the weights ship, and why should you care?

Thinking Machines says it will release the full weights under an Apache 2.0 license next week, after the completion of external safety testing and red-teaming. One tracker lists the week of November 10 to 14, conditional on that safety testing. The date is the lab's promise, not a guarantee, and one outlet noted it could not confirm the release has happened, so watch the lab's own channels rather than assuming a date.

The license matters more than the date. Apache 2.0 is a permissive open license, which means you will be able to download the weights, run the model on your own hardware, fine-tune it without paying Tinker, and build products on it. That is the difference between renting a model and owning one. Mistral Large 4, by comparison, ships under a modified MIT license that requires a separate commercial deal for hosted Model-as-a-Service businesses. Apache 2.0 has no such condition, which is why builders care about this release.

When the weights land, the model also becomes the open alternative to Cognition's SWE-2 Flash, which is built on the first-generation Inkling-Small and is not available as weights at all. If you want to self-host a Thinking Machines model, Inkling 2 Small is the one to wait for.

What are the catches?

The first catch is access. Closed weights on one platform, in beta, with a production waitlist, is a narrow door. If Tinker has an outage, or if your use case needs a different provider, you have no fallback until the weights ship.

The second catch is context. The model card promises up to 1 million tokens, but Tinker serves 64K and 256K today. Long-context work has to wait.

The third catch is evidence. The Intelligence Index score of 31 is real, but it is one index, and the broader benchmark picture is thin. Early coverage also notes the model can be verbose, which means output tokens, and output cost, can run higher than planned on long answers. Budget for it.

The fourth catch is the weights date. "Next week, after safety testing" is a conditional promise. Safety testing can find something, and dates can slip. Do not plan a launch around a weights release that has not happened yet. Use the Tinker API now, and treat the weights as a bonus when they arrive.

Who should try it now?

Try it this week if you build coding tools, agents, or document work and you want frontier-class size at $0.30 per million input tokens. Run your real prompts through the Tinker playground, measure the quality against the model you use today, and count the tokens. The price is low enough that a serious test costs dollars, not hundreds.

Wait if you need a million tokens of context, a production service level agreement, or open weights today. None of those exist yet. Wait if your work depends on a model with a deep independent benchmark record, because Inkling 2 Small does not have one yet. And wait if you need to self-host, because the weights are promised for next week, not available today.

For everyone else, the move is simple: test it now while it is cheap, and watch for the weights. A 594 billion parameter model at $0.30 per million tokens, with Apache 2.0 weights on the way, is the kind of release that changes what small teams can afford to build. The price is verified. The quality is promising. The weights are the part still to prove.

Sources: Thinking Machines Lab, Introducing Inkling 2 Small, November 5, 2026; Tinker models and pricing page, as read November 7, 2026; Artificial Analysis Intelligence Index, as reported November 2026; DepthGrid, Thinking Machines Inkling 2 Small, November 7, 2026; Kie.ai, What Is Inkling 2 Small, November 2026; TechToHeart, Inkling 2 Small, November 2026.

Related reading: AI Basics & Tutorials, A 65B Open Model Just Beat a 501B Giant at Coding, Mistral Large 4 Open Weights Are Out

Frequently Asked Questions

How much does Inkling 2 Small cost?

On Tinker serverless inference, $0.30 per million input tokens, $0.03 per million cached input tokens, and $1.00 per million output tokens. Fine-tuning costs $1.50 per million tokens at 64K context and $3.00 at 256K.

Can I download the weights?

Not yet. Thinking Machines says the full weights ship under an Apache 2.0 license next week, once external safety testing and red-teaming finish. The date is the lab's promise, not a guarantee.

Where can I try Inkling 2 Small today?

Only on Tinker, through the playground or the OpenAI-compatible API. No other platform serves it yet, and the weights are not public.

What context window do you actually get?

The model card lists up to 1 million tokens, but Tinker serves 64K and 256K variants today. The full window is not available by default on the platform yet.

How does it compare with Mistral Large 4?

Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, and Artificial Analysis puts the cost of running its index at $1.18 per task for Inkling against $2.06 for Mistral Large 4. Both figures are index results, not independent evaluations of your workload.

Is Inkling 2 Small open source?

Not yet. It runs as a hosted proprietary model on Tinker today. The Apache 2.0 weights release is promised for next week, after safety testing, and only then does it become open weights.