Thinking Machines Chose Open Weights First

Seventeen months after Thinking Machines Lab publicly introduced itself, the company released its first foundation model on July 15, 2026. The unusual part is not that Inkling is large, multimodal, or trained by people who helped build ChatGPT. The unusual part is that the full weights arrived first.

Inkling is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, a one-million-token context window, and native text, image, and audio input. The weights are on Hugging Face under Apache 2.0. Thinking Machines is explicit that Inkling is not the strongest model available, open or closed.

That concession makes the release more credible. Inkling does not need to win every benchmark for its launch strategy to matter. It needs to be good enough that open weights are a serious product decision rather than a consolation prize. The early evidence clears that bar.

For the current score, rank, pricing, context window, architecture, and complete benchmark ledger, use the Thinking Machines Inkling model profile. This article owns a different question: why a heavily funded frontier lab chose open weights for its first foundation-model release, and whether the evidence makes that choice credible.

The first product was not the model

Thinking Machines was publicly unveiled on February 18, 2025 by Mira Murati, the former chief technology officer of OpenAI. The initial team included researchers and builders with histories at OpenAI, Meta, Mistral, and other frontier labs. The company's own description is broader and more useful than a roster: its people helped create ChatGPT, Character.ai, Mistral models, PyTorch, OpenAI Gym, Fairseq, and Segment Anything.

The résumé pile created expectations that Thinking Machines would emerge with another closed frontier API. Its funding made that expectation louder. The company raised a $2 billion seed round at a $12 billion valuation in 2025 before it had released a foundation model.

Instead, its first public product was Tinker, a platform for fine-tuning existing open models. Thinking Machines then previewed a real-time multimodal interaction system. Inkling comes after those pieces, not before them.

That order says something about the company being built.

If the first deliverable had been a closed chat model, Thinking Machines would have entered the same contest as every other frontier lab: accumulate users, sell tokens, and defend an API margin. Tinker established a different center of gravity. The paid product is the machinery for adapting models. Inkling is the base model that makes that machinery more valuable.

Open weights are therefore not a side project attached to the launch. They complete the product loop. A developer can download the checkpoint, run it through common inference stacks, fine-tune it on Tinker, or buy hosted inference from a partner. Thinking Machines can make the base portable while charging for the difficult part: reliable post-training infrastructure.

Why the release mechanics matter

Inkling is a full-scale sparse transformer trained from scratch on 45 trillion tokens of text, images, audio, and video.

Table 1 Download CSV

Release fact	Inkling
License	Apache 2.0
Architecture	66-layer sparse mixture-of-experts transformer
Parameters	975B total, 41B active
Experts	256 routed, 2 shared; 6 routed experts active per token
Context	Up to 1M tokens
Inputs	Text, images, audio
Output	Text
Checkpoints	BF16 and NVFP4
Hosted customization	Tinker

The architecture is interesting where it departs from the default recipe. Thinking Machines says Inkling interleaves sliding-window and global attention at a five-to-one ratio, uses relative position embeddings instead of RoPE, and adds short convolutions around attention and residual paths. The model supports a controllable reasoning effort setting, allowing a deployment to spend more or fewer tokens on a task.

The word "open" needs a hardware footnote. Thinking Machines says the BF16 checkpoint requires at least 2 TB of aggregate VRAM: eight NVIDIA B300 GPUs or sixteen H200s. The quantized NVFP4 checkpoint still needs at least 600 GB. Inkling is downloadable, inspectable, and deployable outside the company's API, but it is not a model most developers will run under a desk.

The license also deserves precision. Apache 2.0 is permissive, and the weights are available. That does not make the full training dataset, every training decision, or the entire production system open source. "Open weight" is the accurate term.

The web-app result is the sharpest signal

The most useful independent result on launch day is not a multiple-choice science score. It is Inkling's position on Design Arena's Agentic Web Dev leaderboard, where users compare deployed web applications without seeing the model names.

The launch article captured Inkling at 1,257, tied with Claude Opus 4.6 for seventh. The live board has already moved. At publication, Design Arena reports Inkling under its evaluation codename camellia at 1,258 Elo from 1,152 battles. That places it tenth on the benchmark and second among open-weight models, behind GLM-5.2.

Table 2 Download CSV

Live Agentic Web Dev position	Model	Weight access	Elo
1	Claude Sonnet 5	Closed	1,334
4	Claude Opus 4.8	Closed	1,289
5	GLM-5.2	Open	1,280
6	Grok 4.5	Closed	1,278
7	Claude Opus 4.6	Closed	1,261
10	Inkling	Open	1,258
12	Kimi K2.7 Code	Open	1,240

The live movement is not a discrepancy to hide. It is how a preference leaderboard works. Ratings and ranks change as votes arrive. The stable finding is narrower: Inkling's generated applications are being preferred at a rate close to several current closed leaders, and only one open-weight model sits above it on this board.

Design Arena is not a general intelligence test. It evaluates a specific workflow: agents build multi-file React applications from real user prompts, and humans judge the rendered output. The harness observes multi-file edits, tool calls, retries, long-running execution, and failure recovery. That makes it unusually relevant to the product Inkling is supposed to become: a customizable model inside tools, not just an answer engine inside a chat box.

We keep the row display-only: it appears on the Inkling model page but does not change the weighted score. A moving human-preference Elo should not silently become a general capability point.

The broader benchmark table is credible, not dominant

One launch-day reaction compressed the appeal and the caveat into the same post. It is commentary, not benchmark evidence, but the tension is the right one to examine.

Thinking Labs finally released their first big model as part of a model family, including smaller models

it's an MoE with 975B @ 41B parameters, trained on 45 trillion tokens and it's open-weight!

noticeably it reasons over text, images, and audio

but benchmarks don't look…

— Lisan al Gaib (@scaling01), July 15, 2026

Inkling's launch table is wide. At its highest published effort setting, the model scores 97.1% on AIME 2026, 87.9% on GPQA Diamond, 77.6% on SWE-bench Verified, 54.3% on the public SWE-bench Pro split, 63.8% on Terminal-Bench 2.1 with its best harness, 77.1% on BrowseComp with context management, and 74.1% on MCP Atlas.

The pattern is easier to read when the rows are separated by what they prove.

Table 3 Download CSV

Capability	Inkling result	What it supports
AIME 2026	97.1%	Frontier-level competition math, on a saturated row
GPQA Diamond	87.9%	Strong graduate science reasoning
SWE-bench Verified	77.6%	Credible repository issue resolution
SWE-bench Pro Public	54.3%	Useful but not leading difficult software work
Terminal-Bench 2.1	63.8%	Capable terminal agent, behind the best closed systems
BrowseComp	77.1%	Strong web research with context management
MCP Atlas	74.1%	Competitive tool use and coordination
IFBench	79.8%	Strong detailed instruction following
MMMU Pro	73.5%	Competitive multimodal reasoning
CharXiv with Python	82.0%	Strong chart and paper understanding with tools

Some rows place Inkling near current closed leaders. Others do not. In our dataset, Claude Fable 5 holds 95.0% on SWE-bench Verified and 84.3% on Terminal-Bench 2.1, and GPT-5.6 Sol posts 91.9% on the same terminal benchmark. Inkling's 30.0% on text-only Humanity's Last Exam trails GLM-5.2, the strongest open-weight model on that row, as well as the leading closed systems.

That mixed shape is why the BenchLM score is 69, ranking thirtieth among 78 currently eligible models. Inkling has 16 trusted benchmark rows and enough coverage to rank. It does not have sourced multilingual results in the current dataset, and its reasoning coverage does not support a category score. The aggregate rewards the evidence that exists. We are not going to fill the gaps with launch-day enthusiasm.

The model's own launch post uses similarly careful language. Thinking Machines calls Inkling broad and balanced, then states that it is not the strongest overall model. Labs usually bury that sentence under a chart. Here it is part of the positioning.

Open weights change what a first model can do

A closed first model has one distribution path: the company operating it. An open-weight first model can spread through inference providers, research groups, enterprise clusters, framework integrations, fine-tunes, quantizations, and derivative products before the original lab has built every route itself.

That distribution is valuable to a new lab, even a well-funded one.

Thinking Machines does not yet have the consumer reach of ChatGPT, the enterprise footprint of Microsoft and Google, or the developer default status of Anthropic inside coding tools. Publishing weights lets other companies do integration work that would otherwise sit on Thinking Machines' roadmap. The model card already points developers toward SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face deployment paths.

The checkpoint also turns claims about customization into something testable. A company can fine-tune Inkling, inspect the resulting weights, deploy them on its own infrastructure, and keep using the model if its relationship with Thinking Machines changes. That is a different procurement proposition from renting a closed endpoint whose behavior, availability, and terms remain controlled by the vendor.

There is a trade. Thinking Machines gives up some control over distribution and downstream behavior. It also makes safety work harder because fine-tuning can alter the safeguards described in the model card. The company acknowledges that problem and says it is studying how customization changes safety behavior.

The release is still commercially coherent. Open weights generate adoption; Tinker sells customization; hosted partners sell inference. The model is the substrate, not the entire business.

The economics are less romantic

Open weights do not automatically mean lower total cost.

Inkling's active parameter count is efficient relative to its 975B total size, but the memory requirement remains substantial. A team choosing self-hosting needs enough accelerators, an inference stack, utilization high enough to justify the cluster, and engineers who can keep it reliable. The NVFP4 checkpoint makes deployment more practical; 600 GB of aggregate VRAM is still a serious system.

Tinker offers the convenient path. Its launch pricing lists Inkling's 64K tier at $1.87 per million uncached input tokens, $0.374 for cached input, and $4.68 for output during a limited-time 50% discount. The 256K tier doubles those rates. Those are hosted-service economics, not proof that possession of the weights makes inference free.

The practical buyer split looks like this:

• Use hosted Inkling when customization matters but operating a large model does not.

• Use the checkpoint when data control, model independence, or sustained utilization can justify the cluster.

• Use a smaller open model when the workload does not need Inkling's multimodality, context, or capability.

• Keep a stronger closed model in the route for tasks where Inkling's quality gap costs more than the token savings.

That last option is not a failure of the open release. Model routing is where an open-weight model this capable becomes most useful. It does not have to replace the closed frontier. It can take the large middle of the workload while a smaller model handles cheap tasks and a closed flagship handles the hardest tail.

A better test than launch-day rank

The next six months will tell us more than the first benchmark table.

The immediate questions are operational. Can inference providers serve the 41B-active model at attractive latency? Does NVFP4 preserve the capabilities buyers care about? Do fine-tunes move specialized performance without damaging the model's broad base? Can teams reproduce the advertised long-context behavior outside Tinker? Does the open ecosystem build useful derivatives, or does the model remain too large for wide experimentation?

There are also evidence gaps we cannot close from here. Independent multilingual results would make the generalist claim easier to assess. More benchmark-owned coding and agent evaluations would reduce dependence on provider-run harnesses. Reproducible cost-per-task measurements would show whether controllable effort produces savings outside Thinking Machines' own setup.

Design Arena will keep moving. That is useful, but its narrowness should remain visible. A model that makes preferred React apps is not therefore the best model for science, legal analysis, support automation, or multilingual retrieval.

The strongest possible outcome for Inkling is not holding seventh, tenth, or any other launch-week rank. It is becoming a base that other people improve in ways Thinking Machines did not plan.

That is the wager embedded in the weights.

The strategic result

Thinking Machines spent its first seventeen public months talking about collaboration, customization, shared science, and infrastructure. Inkling could have exposed those ideas as branding wrapped around another closed API. Instead, the company released the checkpoint that makes the claims falsifiable.

The result is not an overall frontier leader. It is more interesting than a weak open model and more limited than the best closed systems. Its benchmark profile is broad enough to support real deployment, its Design Arena result is close enough to closed leaders to earn attention, and its hardware demands are high enough to keep the economics honest.

For a lab built by people from OpenAI and funded before it had a model, choosing open weights for the first foundation release is a meaningful signal. Thinking Machines is not merely trying to rent intelligence. It wants to sell the tools used to reshape it.

Inkling is the first real test of whether those two businesses can be the same company.