Thinking Machines Inkling: 975B Open-Weight Model

Thinking Machines Inkling is the first model from Mira Murati's Thinking Machines Lab, released on July 15, 2026 under an Apache 2.0 licence. It is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters per forward pass, trained on 45 trillion tokens spanning text, images, audio and video, with a one-million-token context window. The weights are freely downloadable on Hugging Face, and the company is explicitly not selling metered API access to it.

The unusual part is not the size. It is the stated goal. Thinking Machines has been openly uninterested in topping leaderboards, arguing that most enterprises do not need the smartest general model — they need one they can make their own. Inkling is engineered as a base to be customized, and the company's fine-tuning platform, Tinker, is the actual product.

Below: the full specifications, what "built for customization" means in practice, how it stacks up against the other open-weight options in 2026, and the honest assessment of who should and should not pick it up.

Key Takeaways

• Inkling launched July 15, 2026 under Apache 2.0 — genuinely permissive, not a source-available lookalike licence.
• 975B total parameters, 41B active, trained on 45T tokens across text, image, audio and video, with a 1M-token context window.
• A companion model, Inkling-Small, is a 276B-total / 12B-active MoE for cheaper, faster inference.
• Weights are on Hugging Face; hosted inference is live on Together AI, Fireworks, Modal, Databricks and Baseten.
• Thinking Machines is not monetizing Inkling through per-token API pricing — the business is Tinker, its fine-tuning platform.

What is Thinking Machines Inkling?

Thinking Machines Inkling is an open-weight, natively multimodal foundation model built as a customization base rather than a finished product. It uses a mixture-of-experts architecture: 975 billion parameters exist in the model, but only about 41 billion activate for any given token, which is what keeps inference cost in a range a normal company can actually pay.

Native multimodality matters here in a specific way. Many "multimodal" models are text models with vision bolted on afterwards. Inkling was pretrained across text, images, audio and video from the start, which generally produces better cross-modal grounding and — more relevant for fine-tuning — a representation space that adapts to mixed-modality domains without fighting the base model.

TechCrunch framed the launch as a bet against one-size-fits-all AI, and that is the accurate reading. Axios noted that this is the lab's first in-house model since Murati, OpenAI's CTO until late 2024, founded the company.

Inkling specifications at a glance

The 41B active parameter count is the number to plan around. It determines memory bandwidth and therefore serving cost — a 975B model with 41B active behaves, at inference time, much more like a 41B dense model than like a trillion-parameter one, provided your serving stack handles expert routing efficiently. Storage and loading, however, are governed by the full 975B, which is the practical barrier for self-hosting.

Compare that to the alternatives: DeepSeek's V4 Flash runs 284B total / 13B active and is API-first, while Inkling is roughly three times the size and weights-first. Different tools for different jobs — see our DeepSeek V4 Flash 0731 breakdown if raw cost-per-token is your constraint.

How do you actually use Inkling?

Three paths, in ascending order of effort.

1. Hosted inference. The fastest route. Inkling is live on Together AI, Fireworks, Modal, Databricks and Baseten. Most expose an OpenAI-compatible endpoint:

curl https://api.together.xyz/v1/chat/completions \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/inkling",
    "messages": [{"role": "user", "content": "Summarise this contract clause."}]
  }'

2. Fine-tuning on Tinker. This is the intended path and the company's business model. You bring a domain dataset, Tinker handles the training infrastructure, and you get a customized model back.

3. Self-hosting the weights. Fully possible under Apache 2.0 and genuinely demanding. Downloading 975B parameters is a serious storage and networking exercise before you serve a single token, and you will want a server that handles MoE routing well:

# Pull the weights (expect this to take a while and a lot of disk)
huggingface-cli download thinking-machines/inkling --local-dir ./inkling

# Serve with an MoE-aware inference engine
vllm serve ./inkling \
  --tensor-parallel-size 8 \
  --max-model-len 131072

Note the reduced --max-model-len. The model supports a million tokens; your KV cache budget probably does not. Start well below the maximum and raise it only when you have measured memory headroom. Our vLLM vs Ollama comparison covers why the serving engine choice dominates throughput for models in this class, and how to run LLMs locally covers the smaller-scale version of the same problem.

Is Inkling better than other open-weight models?

Not on general benchmarks, and Thinking Machines says so directly. The company has been unusually candid that Inkling is not the strongest model available, and that its advantage shows up after customization rather than before it.

That is an honest and somewhat risky position. Benchmark scores are how models get adopted, and declining to compete on them means declining most of the free attention in the market. The bet is that a fine-tuned 41B-active model on your domain beats a general frontier model on your domain — which, in our experience with production fine-tuning, is frequently true for narrow, well-specified tasks and frequently false for open-ended reasoning.

If you have never fine-tuned a model before, start smaller and cheaper to learn the workflow — our guide to Unsloth and faster LLM fine-tuning walks through the mechanics on hardware you probably already have.

Why the Apache 2.0 licence is a bigger deal than the parameter count

"Open weights" has become one of the least trustworthy phrases in AI. Several widely used models ship under bespoke licences that cap monthly active users, forbid competing products, restrict entire application categories, or reserve the right to change terms later. Legally, those are not open source — they are conditional distribution agreements, and they create exactly the kind of ambiguity that makes company lawyers say no.

Apache 2.0 removes that entire conversation. It is an OSI-approved licence with no usage caps, no revenue thresholds, no field-of-use restrictions, and an explicit patent grant. In practice that last clause matters more than developers expect: the patent grant protects you from the licensor asserting patents against your use of the model, which is not something a custom "community licence" reliably gives you.

The practical consequences for a team evaluating Inkling:

• You can build a competing product on it. Nothing in the licence prevents you from shipping a commercial AI service built on Inkling, even one that competes with Thinking Machines.
• You can fork and redistribute. Fine-tune it, rename the artifact, publish the derivative, sell it. Attribution and the licence text are the only obligations.
• The terms cannot be revoked retroactively. The version you downloaded stays licensed under the terms you got it under.
• Procurement review is short. Apache 2.0 is already on virtually every enterprise's pre-approved licence list, which removes weeks of legal review that a bespoke licence would trigger.

For anyone who has watched a genuinely promising model get rejected in procurement over a licence clause nobody could interpret, this is not a footnote. For a broader look at what a healthy permissive ecosystem enables, see our roundup of open source AI developer tools.

The angle worth noticing: giving away the model to sell the workflow

Most coverage filed this as "another open-weight release." The more interesting story is the business model, because it inverts the standard AI economics.

Every major lab currently monetizes inference: you pay per token, forever, and the model is the product. Thinking Machines is giving the model away under Apache 2.0 and monetizing customization. That means the company profits when you succeed at making the model yours — the exact opposite incentive from a per-token vendor, who profits when your prompts get longer and your usage grows.

There is a real strategic logic underneath it. Inference is commoditizing fast; the price of a million tokens has collapsed by roughly an order of magnitude in eighteen months, and it will keep falling. Fine-tuning workflows, evaluation harnesses and the operational knowledge of how to adapt a base model to a domain are much stickier. If Murati is right that the durable value sits in customization rather than raw capability, this is a well-timed land grab. If she is wrong, Thinking Machines has given away a very expensive asset for the privilege of selling a service anyone can build.

Either way, developers win: a permissively licensed 975B multimodal model with a million-token context is a genuine gift to the open ecosystem, regardless of whether the business thesis holds.

Frequently Asked Questions

What is Thinking Machines Inkling? Inkling is the first foundation model from Thinking Machines Lab, released July 15, 2026 under Apache 2.0. It is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, trained on 45 trillion tokens across text, images, audio and video, with a one-million-token context window.

Is Inkling really free to use commercially? Yes. It ships under Apache 2.0, one of the most permissive open-source licences, with no usage caps, no revenue thresholds and no field-of-use restrictions. You can download, modify, deploy and sell products built on it.

How big is Inkling-Small? Inkling-Small is a mixture-of-experts model with 276 billion total parameters and 12 billion active, aimed at teams needing faster and cheaper inference. It launched in preview alongside the flagship.

Where can I run Inkling without my own GPUs? Hosted inference is available on Together AI, Fireworks, Modal, Databricks and Baseten. For fine-tuning, Thinking Machines offers Tinker, its own customization platform.

Is Inkling better than DeepSeek or Llama? Not on general benchmarks — Thinking Machines has said plainly that Inkling is not the strongest model available. Its case rests on being an unusually good base for fine-tuning, particularly for multimodal domains, thanks to native pretraining across four modalities.

What hardware do I need to self-host Inkling? Enough VRAM and storage for 975 billion parameters plus KV cache, which in practice means a multi-GPU node with tensor parallelism and an MoE-aware serving engine. Most teams should use hosted inference or Inkling-Small unless self-hosting is a hard requirement.

The verdict

Inkling is the most interesting open-weight release of 2026, and the reason is not its benchmark position — it is the licence and the business model behind it. Apache 2.0 on a 975B natively multimodal model with a million-token context is a serious contribution to the commons, and doing it while explicitly declining to compete on leaderboards is a statement about where the industry's value is heading.

Our recommendation: if you have a narrow, data-rich domain and the appetite to fine-tune, evaluate Inkling seriously — this is the best-positioned base model for that job right now. If you need the strongest general reasoning available today, or if your workload is high-volume generic inference where cost per token dominates, look elsewhere; Inkling is not trying to win that fight.

Ready to actually try customizing a model? Start with our practical walkthrough of Unsloth and faster LLM fine-tuning, then scale the same workflow up.

Every other lab is selling you intelligence by the token. Thinking Machines just gave away the intelligence and offered to sell you the screwdriver.