TL;DR
Thinking Machines Lab released the full weights for Inkling, its first foundation model, under Apache 2.0 before offering a closed API. The release gives developers greater control over deployment and modification, but the flagship’s hardware demands, unpublished training data and a reported use policy limit what ownership means in practice.
Thinking Machines Lab, founded by former OpenAI technology chief Mira Murati, has released the full weights for its first foundation model, Inkling, on Hugging Face under Apache 2.0. Publishing the weights before a closed API gives organizations direct control over deployment and modification, although the model’s large hardware requirements mean most developers cannot run the flagship locally.
Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters. According to the laboratory, it has a 1-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 release includes BF16 and NVFP4 checkpoints and day-one support for Transformers, vLLM, SGLang, llama.cpp, TokenSpeed and Unsloth. Under the stated Apache 2.0 license, developers can download, modify and commercialize the weights. The training data and full training pipeline have not been published, meaning open weights do not amount to a fully open-source system.
The laboratory also acknowledged that Inkling is not the strongest available model, whether compared with closed or open alternatives. Vendor-published results show strengths on AIME 2026, GPQA Diamond, VoiceBench and adversarial testing, while results supplied with the launch place it behind models including GLM-5.2 on several coding and reasoning tests. Some scores use a prerelease checkpoint or data from Artificial Analysis and have not received independent 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.
Model Ownership Moves Upfront
Releasing the weights at launch changes what customers receive. An API user rents access under a provider’s pricing, availability and product rules; a weight holder can run the model on chosen infrastructure, retain modified versions and build specialized systems without sending every request to the original developer. That can support data control, operational continuity and custom fine-tuning.
The release also indicates that Thinking Machines Lab sees distribution and developer control as competitive assets, rather than treating open weights as a delayed version of a commercial service. Its public admission that Inkling does not lead every benchmark shifts attention toward deployment economics and control, including a configurable reasoning-effort setting from 0.2 to 0.99 that trades additional computation for higher performance.
Those benefits are unevenly distributed. The source material estimates that BF16 deployment requires at least 2 terabytes of aggregate graphics memory, while NVFP4 still needs about 600 gigabytes. For many teams, the practical choice remains a hosted service, a compressed version or the planned Inkling-Small model, rather than direct operation of the flagship.
Ollama: Run the AI Models You Choose on Your Own PC
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Open Access Stops at Training
Open-weight releases allow developers to inspect and alter learned parameters, but they do not automatically disclose the material or methods used to create them. Thinking Machines Lab published Inkling’s checkpoints and deployment integrations, while withholding its training dataset and full pipeline, a boundary that limits outside auditing and reproduction.
The company describes Inkling as a Western alternative in an open-model market where Chinese-developed systems remain strong competitors. The supplied benchmark comparisons place GLM-5.2 ahead on several agentic and reasoning tasks and Kimi K2.6 ahead on some multimodal work. The source also reports that Kimi K2.5 generated synthetic data used during Inkling’s post-training, illustrating how development methods cross competitive boundaries.
A smaller preview, Inkling-Small, has 276 billion total parameters and 12 billion active parameters. Thinking Machines Lab says it matches or exceeds the flagship on several tests, but its full weights are due only after testing. That model may prove more accessible to organizations without data-center-scale hardware.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab, in its Inkling announcement
License and Benchmarks Need Checking
It is not yet clear whether every commercial use permitted by Apache 2.0 is also subject to a separate Model Acceptable Use Policy. The source material reports restrictions covering surveillance, deception and fully automated decisions affecting rights, but says that policy was not independently verified. Organizations working in geospatial analysis, public safety or other regulated areas will need to examine the model card and repository terms directly.
The model’s comparative performance also remains unsettled. Most cited results are vendor-published benchmarks, some involve a prerelease checkpoint, and independent researchers have not yet reproduced them. Real-world cost, latency, reliability and multimodal performance may differ from the supplied scores, while the reported efficiency of the adjustable reasoning setting still needs testing across varied workloads.
Independent Tests and Smaller Weights
Researchers and prospective users are expected to test Inkling’s published checkpoints, verify the governing terms and compare its performance with GLM-5.2, Kimi K2.6 and closed systems on their own workloads. The results will show whether the model’s control-versus-compute tradeoff is commercially useful beyond benchmark reports.
Attention will also turn to the release of Inkling-Small’s full weights after testing. Its lower active-parameter count could make the laboratory’s open-first strategy relevant to more developers, although hardware needs, final licensing terms and the release date remain unconfirmed.
Key Questions
What are AI model weights?
Model weights are the numerical values learned during training that govern how an AI system processes inputs and produces outputs. Access to them allows developers to host, modify and fine-tune a model instead of relying solely on its creator’s API.
Is Inkling fully open source?
No. Thinking Machines Lab released the model weights under Apache 2.0, but it did not publish the complete training dataset or pipeline. The release is best described as open weight rather than fully reproducible.
Can Inkling run on a normal workstation?
The flagship is unlikely to run at full quality on ordinary equipment. The supplied estimates call for at least 2 terabytes of graphics memory for BF16 or about 600 gigabytes for NVFP4, placing standard deployment in data-center territory.
Does Inkling lead the major AI benchmarks?
No overall lead has been established. Vendor-published results show strong performance in mathematics, scientific reasoning, audio and adversarial tests, but Inkling trails some rivals on coding, agentic and multimodal evaluations. Independent replication is pending.
Why does releasing the weights first matter?
The order gives customers direct possession of the model parameters from launch rather than making access dependent on a closed service. That can reduce provider dependence and permit private deployment, but hardware costs and possible use restrictions still constrain practical control.
Source: Thorsten Meyer AI
