TL;DR
Thinking Machines Lab released its first foundation model, Inkling, with downloadable weights under Apache 2.0 and immediate support from major deployment tools. The open-first approach offers users more control, although the model requires costly hardware and its benchmark claims still need independent testing.
Thinking Machines Lab, founded by former OpenAI technology chief Mira Murati, released its first foundation model, Inkling, on July 15 with full downloadable weights under Apache 2.0 before offering a closed API. The release matters because it gives companies direct control over deployment and modification from day one, even as the lab acknowledges that Inkling is not the strongest available model.
Inkling is a mixture-of-experts model with 975 billion total parameters and 41 billion active parameters. Thinking Machines says it was pretrained on 45 trillion tokens, supports a one-million-token context window, and can receive text, images and audio while producing text.
The lab published BF16 and NVFP4 checkpoints on Hugging Face and provided immediate compatibility with Transformers, vLLM, SGLang and llama.cpp, among other tools. Under Apache 2.0, developers can generally download, modify and use the model commercially. The source material reports a separate acceptable-use policy, however, and its legal effect has not been independently verified.
Thinking Machines reports strong results on AIME 2026, GPQA Diamond and VoiceBench, while placing behind rivals on several software-engineering and agent benchmarks. The lab also offers a 0.2-to-0.99 reasoning-effort control intended to trade computing cost and latency against performance. These figures are vendor-published results, some involving a prerelease checkpoint, and await outside replication.
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.
Open-First Release Changes the Bet
Most frontier-model providers sell access through hosted services while retaining control over the underlying models. Inkling reverses that sequence: the weights arrived first, giving qualified users the ability to run the model on their own infrastructure, modify it and avoid dependence on a single hosted API.
That approach could appeal to organizations concerned about data control, service access or long-term costs. It also shifts attention from a single leaderboard score to the economics of operating a model at different reasoning settings. Yet ownership does not mean easy access: the full model remains beyond the hardware budgets of many developers.
Accelerate Everything with Tensor Cores: A Developer’s Guide to High-Performance AI, Efficient Training, and Scalable Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Inside Inkling’s Technical Release
Thinking Machines Lab is about 17 months old and employs several researchers who previously worked on ChatGPT. Inkling uses a 66-layer decoder-only design that routes tokens among 256 experts, with six selected experts and two shared experts active during processing.
The launch also included a preview of Inkling-Small, a 276-billion-parameter model with 12 billion active parameters. Thinking Machines says the smaller version matches or exceeds the flagship on some tests, but its complete weights have not yet been released. For local and lower-cost deployment, that model may prove more accessible than the flagship.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab’s launch announcement
License and Performance Questions Persist
It remains unclear whether a reported Model Acceptable Use Policy adds binding restrictions beyond Apache 2.0 for the original parameters or modified versions. The reported limits cover surveillance, deception and automated decisions affecting rights. Organizations in public safety, intelligence or geospatial work would need to inspect the controlling documents before adopting the model.
Independent evaluators have not yet confirmed the benchmark scores or efficiency claims. Several competing models, including GLM-5.2 and Kimi K2.6, reportedly remain ahead on some reasoning, coding and multimodal tasks. Inkling’s real-world cost curve also depends on workloads, quantization and hardware configuration.
Independent Tests and Smaller Weights Awaited
Researchers and prospective customers will next compare Inkling against GLM-5.2, Kimi K2.6 and closed models on production workloads. Attention will also turn to clarification of the acceptable-use terms, independent replication of benchmark results and the promised release of Inkling-Small’s full weights. Those developments will show whether the open-first strategy produces practical gains beyond the launch itself.
Key Questions
What did Thinking Machines Lab release?
The company released Inkling, its first foundation model, including full BF16 and NVFP4 weights for the 975-billion-parameter mixture-of-experts system.
Is Inkling open source?
Its weights are available under Apache 2.0, but the training data and full training pipeline have not been published. It is more precise to call Inkling an open-weight model.
Can Inkling run on a workstation?
Not in its standard forms. The source estimates that BF16 needs at least 2 terabytes of aggregate VRAM, while NVFP4 still requires about 600 gigabytes. Quantized versions may reduce that requirement with possible quality losses.
Is Inkling the leading open model?
Thinking Machines says it is not the strongest model overall. Its reported results are competitive in mathematics, science, audio and adversarial tests, but rivals remain ahead on several coding, agent and multimodal benchmarks.
Why does releasing the weights first matter?
It lets organizations host, modify and study the model directly without waiting for a future open release. That can provide more control than renting access through a provider-managed API.
Source: Thorsten Meyer AI
