Mira Murati's new lab shipped its first model yesterday, and the most interesting line in the announcement is not a benchmark score. It is this: "Inkling is not the strongest overall model available today, open or closed." A lab run by the former CTO of OpenAI, releasing its first model after 18 months of work, chose to lead with that. That is a positioning statement, and it is worth taking seriously.
What Inkling actually is
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active during inference. It supports context windows of up to one million tokens and was pretrained from scratch on 45 trillion tokens spanning text, images, audio, and video.
The model processes text, image, and audio inputs natively, with capabilities covering reasoning, coding, tool use, instruction following, visual analysis, speech transcription, and long-form audio understanding.
The architecture is a 66-layer decoder-only transformer with a sparse MoE feed-forward backbone: each token is routed to 6 of 256 experts, plus 2 shared experts active on every token. That routing is why you get 975B total parameters but only draw on 41B per forward pass - the model is large enough to encode broad knowledge, cheap enough to serve at scale.
A second model, Inkling-Small, was previewed with 276 billion total parameters and 12 billion active. Thinking Machines says the smaller model matches or exceeds the large version on several internal evaluations while reducing cost and latency. Small's weights are not yet released.
Where it lands against the open-weight field
Inkling scored 77.6% on SWE-bench Verified, 97.1% on AIME 2026, 87.2% on GPQA Diamond, and 73.5% on MMMU Pro. Those are strong numbers in absolute terms, but the open-weight field is crowded right now.
| Model | SWE-bench Verified | GPQA Diamond | MCP Atlas | Active params |
|---|---|---|---|---|
| GLM-5.2 | ~81%* | 91.2% | - | 40B / 744B |
| DeepSeek V4 Pro | 80.6% | - | - | - |
| Kimi K2.6 | - | 91.1% | - | - |
| Inkling | 77.6% | 87.2% | 74.1% | 41B / 975B |
| Nemotron 3 Ultra | 70.7% | - | 44.7% | - |
*GLM-5.2 SWE-bench score is on SWE-bench Pro, not Verified - not directly comparable.
DeepSeek maintains an edge in strict coding and factuality, beating Inkling on SWE-bench Verified (80.6% vs. 77.6%) and SimpleQA Verified (57.0% vs. 43.9%).
Inkling posts 97.1% on AIME 2026 and 77.6% on SWE-bench Verified, beating Nemotron's 94.2% and 70.7%, and significantly leads in agentic workflows, scoring 74.1% on MCP Atlas against Nemotron's 44.7%.
The efficiency story is the more interesting number. The company says Inkling can match Nemotron 3 Ultra on Terminal-Bench while using roughly one-third as many generated tokens. That is an efficiency claim from the launch evaluation, not an independent cost study, but it points to a practical advantage: a model used millions of times may be valuable because of the shape of its performance curve, not merely its maximum score.
The real bet: Tinker, not the weights
Rather than monetize inference tokens on a proprietary closed model, Thinking Machines gives the weights away and sells the customization toolchain - Tinker - capturing the adaptation layer instead of the raw intelligence layer. This is a picks-and-shovels position: the value accrues not to whoever runs the most impressive general model, but to whoever owns the workflow by which organizations make a model their own.
That theory already has one concrete data point. Working with Bridgewater Associates, researchers used Tinker to fine-tune an open model on specialised financial data and produced a lightweight system that scored 84.7% on leading financial reasoning benchmarks, beating the most advanced proprietary alternatives at under 10% of the cost.
On one benchmark, the company says, Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to hit the same coding performance. If that holds across real workloads, the economics of running a fine-tuned Inkling for a narrow domain task start to look meaningfully different from routing every query through a frontier API.
The honest counterargument: the fine-tuning and customization market is already competitive - Hugging Face, Replicate, major cloud providers, and the frontier labs themselves all offer fine-tuning surfaces - and owning the adaptation layer is only defensible if the base model earns developer loyalty first.
Beagle in action#engineering, Thursday afternoonInfrastructure reality before you download
The BF16 checkpoint requires a GPU cluster with at least 2 TB of aggregated VRAM.
The NVFP4 checkpoint offers a quantized alternative that reduces the aggregated VRAM requirement to at least 600 GB. Neither option is a laptop or a single H100. For most teams, this is an API-first model until inference providers scale it out.
API access is available through Together, Fireworks, Modal, Databricks, and Baseten, with inference support across SGLang, vLLM, TokenSpeed, llama.cpp, and Hugging Face Transformers. That is a broad enough set of inference surfaces that you can start experimenting today without owning the hardware - a teammate like Beagle can route tool calls through any of those providers without a model swap.
Inkling itself was built from scratch in under nine months, against the multi-year timelines its rivals run. Whether a nine-month model family can keep pace with labs running multi-year roadmaps is an open question. Thinking Machines pre-trained Inkling from scratch, but used other open-weight models - including Moonshot AI's Kimi K2.5 - to help generate some early post-training data before large-scale reinforcement learning took over. The next model, the company insists, will use fully self-contained post-training instead.
One thing the release gets right: safety evaluation was done before launch rather than after. The team concluded that Inkling did not present risk of material uplift beyond what is already available in the open-weight ecosystem. The residual risks - specifically, Inkling's occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics - are consistent with any open-weight model, and are best addressed with defense-in-depth rather than relying on the model's refusals alone.
Inkling open-weight model: common questions
What is Inkling and who made it?
Thinking Machines Lab, an AI research and product company founded by former OpenAI CTO Mira Murati, released Inkling on July 15, 2026. It extends the company's Tinker customization platform and is intended to serve as the reasoning layer behind its previously previewed real-time voice and vision systems.
How does Inkling compare to GLM-5.2 and DeepSeek V4 Pro?
Inkling trails both on peak agentic coding: DeepSeek V4 Pro beats Inkling on SWE-bench Verified (80.6% vs. 77.6%).
GLM-5.2 scores 91.2% on GPQA Diamond , ahead of Inkling's 87.2%. Inkling's differentiation is multimodal breadth, controllable thinking effort, and a polished fine-tuning pathway - not raw benchmark rank.
Can I run Inkling locally?
Only with substantial hardware. The BF16 checkpoint requires at least 2 TB of aggregate GPU memory, while the NVFP4 checkpoint lowers that requirement to at least 600 GB. For teams without that infrastructure, the practical path is the API via Together AI, Fireworks, or Baseten.
What is the Tinker platform?
Inkling is available for fine-tuning on Tinker today.
Tinker offers Inkling with 64K and 256K context options and a temporary 50% discount. The platform is how Thinking Machines plans to generate revenue - weights are free, adaptation tooling is the product.
Is Inkling safe to deploy in production?
Common downstream moderation tools, such as Llama Guard, are compatible with Inkling and can be layered around the model to catch jailbreak attempts, filter unsafe outputs, and enforce use-case-specific policies. Thinking Machines encourages treating input/output classification as part of your deployment stack, especially for consumer-facing or high-traffic applications where adversarial prompting is more likely.