TL;DR
Thinking Machines Lab released its first foundation model, Inkling, with full weights under Apache 2.0 and immediate support for major deployment tools. The model’s size limits local use, its benchmark claims await independent testing, and a reported separate use policy still needs verification.
Thinking Machines Lab, the 17-month-old company founded by former OpenAI technology chief Mira Murati, released its first foundation model, Inkling, on July 15 with full weights under Apache 2.0 before introducing a closed API. The order of release gives organizations immediate control over deployment and modification, although the lab acknowledges that Inkling is not the strongest model currently available.
Inkling is a 975-billion-parameter mixture-of-experts model that activates 41 billion parameters for each token. Thinking Machines Lab says it has a one-million-token context window and was pretrained on 45 trillion tokens spanning text, images, audio and video. It accepts text, image and audio inputs and produces text.
The lab published BF16 and NVFP4 checkpoints through Hugging Face, accompanied by day-one support for Transformers, vLLM, SGLang and llama.cpp. Apache 2.0 generally permits downloading, modifying and commercially distributing the software, making the release more permissive than models offered only through hosted services or narrower source-available licenses.
The release does not include the training data or full training pipeline, meaning Inkling is more accurately described as open-weight than fully open-source. Thinking Machines Lab also previewed Inkling-Small, a 276-billion-parameter model with 12 billion active parameters, but its complete weights are expected only after testing is finished.
The weights came first: what Inkling actually signals
Mira Murati’s lab shipped its first foundation model — and the model isn’t the story. The order of operations is: full weights, Apache 2.0, day one, before any closed API. Plus a rare concession — the lab says it’s not the strongest model available, open or closed.
- AIME 2026 97.1%
- GPQA Diamond 87.2%
- MCP Atlas (Nemotron 44.7%) 74.1%
- VoiceBench · open-weight audio frontier 91.4%
- FORTRESS adversarial · best open 78.0%
- ForecastBench · calibration 61.1
- HLE text-only (GLM-5.2 40.1%) 29.7%
- SWE-bench Pro (GLM-5.2 62.1%) 54.3%
- Terminal-Bench 2.1 (GLM-5.2 82.7%) 63.8%
- SWE-bench Verified (Fable 5 95.0%) 77.6%
- Design Arena · 2nd open, behind GLM-5.2 ~10th
A 0.2 → 0.99 effort setting trades reasoning tokens against cost & latency, so you get a curve, not a point. On Terminal-Bench 2.1 it reportedly matches Nemotron 3 Ultra at ~⅓ the tokens. Peak score is a vanity metric when you serve millions of calls; the cost curve is what ships. (Bonus: its chain of thought compressed on its own during RL — nobody rewarded it; efficiency did.)
Pitched as the Western alternative to Chinese open weights (censorship-resistance training is the differentiator). But GLM-5.2 still wins on agentic/reasoning and Kimi K2.6 often on multimodal: best American open model, second in the open field. The irony — post-training was bootstrapped on synthetic data from Kimi K2.5.
BF16 needs ≥2 TB aggregate VRAM (8× B300 / 16× H200). NVFP4 still needs ≥600 GB. Not a workstation model — a 512 GB fleet falls just short. “Open” ≠ “runnable.” Mitigations: 1-bit GGUFs (~74% acc.), hosted eval routes, and Inkling-Small (12B active) — the release local-first builders actually want.
Open weights used to be a consolation prize. Inkling is a strategic open release — Apache 2.0, natively multimodal, honestly marketed, published complete on day one, optimized for deployment rather than headlines (the model isn’t the product; the fine-tuning platform is). It doesn’t need to win every benchmark for that to matter. The frontier is learning that owning the base beats renting the API — arriving now from the inside. For the sovereignty buyer: ① a real Western hedge against being switched off · ② verify the use policy before you build · ③ check the VRAM, then benchmark vs GLM-5.2 & Kimi K2.6 on your task.
Model Ownership Takes Priority
The release gives companies, researchers and public institutions a Western-developed foundation model that can be hosted and modified without depending entirely on a vendor-controlled API. That can reduce exposure to price changes, service restrictions and provider shutdowns, while allowing organizations to keep sensitive workloads on infrastructure they control.
Inkling also exposes an adjustable reasoning-effort setting ranging from 0.2 to 0.99. According to the lab’s reported results, users can trade reasoning tokens for lower cost and faster responses rather than treating benchmark performance as a fixed point. For high-volume deployments, the cost-and-latency curve may carry more weight than a model’s highest possible score.
Yet ownership does not equal easy access. The BF16 version reportedly requires at least two terabytes of aggregate graphics memory, while NVFP4 still needs about 600 gigabytes. Those requirements place the flagship beyond normal workstations and many small computing clusters.
AI model deployment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
An Open-Weight Market Contest
Open-weight releases have become a competitive front between American laboratories and developers of Chinese models such as GLM-5.2 and Kimi K2.6. Inkling is positioned as a Western alternative with censorship-resistance training, but vendor-published comparisons indicate that GLM-5.2 remains ahead on several reasoning and agentic tasks, while Kimi K2.6 can perform better on some multimodal work.
Thinking Machines Lab reports scores of 97.1% on AIME 2026, 87.2% on GPQA Diamond and 91.4% on VoiceBench. It also reports weaker results on several software and terminal tasks, including 54.3% on SWE-bench Pro and 63.8% on Terminal-Bench 2.1. Some comparisons used a prerelease checkpoint, and the published results have not yet received broad independent replication.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab
foundation AI model weights
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License Limits Need Verification
It is not yet clear whether Apache 2.0 represents the complete set of usage conditions. A separate Model Acceptable Use Policy has been reported as applying to the model’s parameters and modified versions, with restrictions involving surveillance, deception and fully automated decisions affecting rights. That policy was not independently verified in the supplied material.
The distinction could affect organizations working in geospatial analysis, public safety or intelligence-related fields. Prospective users will need to examine the current model card and repository terms rather than relying on the Apache label alone. Inkling’s real-world performance and operating costs also remain unsettled pending independent testing on production workloads.
large language model API access
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Smaller Weights
Researchers and prospective adopters are expected to test Inkling against GLM-5.2, Kimi K2.6 and hosted commercial models using their own workloads. Testing will need to measure accuracy alongside token use, latency, memory demands and serving costs.
The next major release milestone is the publication of Inkling-Small’s full weights after testing. Its 12 billion active parameters could make it more relevant to smaller operators than the flagship, though its hardware requirements and final license conditions have not been confirmed. Users making medical or other high-impact decisions should consult qualified professionals and applicable regulators before relying on model output.
AI model training datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Thinking Machines Lab release?
The company released Inkling’s full BF16 and NVFP4 model weights on Hugging Face under Apache 2.0, together with support for several common inference frameworks.
Is Inkling fully open-source?
No. The weights are available, but the training data and complete training pipeline were not published. A reported additional use policy also requires verification.
Can Inkling run on a local workstation?
Not in its standard flagship formats. BF16 reportedly needs at least two terabytes of graphics memory, and NVFP4 requires about 600 gigabytes. Lower-bit community formats may reduce memory use but can affect accuracy.
Does Inkling lead current model benchmarks?
No. Thinking Machines Lab says Inkling is not the strongest available model. Its published scores show strong mathematics, audio and calibration results but weaker performance on some coding and agentic benchmarks.
Why does the release order matter?
Publishing the weights before a closed API gives users immediate control over hosting, modification and commercial deployment. It frames model ownership, rather than exclusive service access, as the release’s main value.
Source: Thorsten Meyer AI
