TL;DR
Thinking Machines Lab released the full weights for its first foundation model, Inkling, under Apache 2.0 before offering a closed API. The move points toward an AI market focused on model ownership and deployment control, although Inkling remains expensive to run and its performance claims await independent testing.
Thinking Machines Lab, the 17-month-old company founded by former OpenAI technology chief Mira Murati, has released the full weights for its first foundation model, Inkling, under an Apache 2.0 license before introducing a closed API. The order of release matters because it gives qualified users immediate control over downloading, modifying and commercially deploying the model, while the lab openly acknowledges that Inkling does not lead every benchmark.
Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters for each token. According to specifications published with the release, it supports a 1-million-token context window and was pretrained on 45 trillion tokens drawn from text, images, audio and video.
The model accepts text, images and audio and produces text. Thinking Machines says its multimodal components were trained together rather than added later as adapters. The released checkpoints include BF16 and NVFP4 versions on Hugging Face, with launch-day support for Transformers, vLLM, SGLang and llama.cpp.
Vendor-published results place Inkling near the front of several tests, including reported scores of 97.1% on AIME 2026, 87.2% on GPQA Diamond and 91.4% on FORTRESS. It trails other models on several coding and agent tasks, including SWE-bench Pro and Terminal-Bench 2.1. Some results used a prerelease checkpoint, and independent replication is not yet available.
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 Release Reshapes Model Ownership
Releasing the weights first makes ownership and deployment control part of Inkling’s core offer, rather than a later concession. Organizations able to host the model can modify it, fine-tune it for private data and reduce their dependence on an API provider that may change pricing, access or usage limits.
The model also offers a reported 0.2-to-0.99 reasoning-effort setting, allowing operators to trade additional reasoning tokens for cost and response time. Thinking Machines reports that Inkling can match Nemotron 3 Ultra on Terminal-Bench 2.1 while using about one-third as many tokens. That claim could matter to high-volume operators, but it still requires outside testing under production conditions.
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Inkling Joins the Open-Weight Race
Open-weight models have become an alternative to systems available only through hosted interfaces. Inkling enters that market as a US-developed multimodal model competing with Chinese releases including GLM-5.2 and Kimi K2.6, which retain reported advantages on some reasoning, agent and multimodal tests.
Thinking Machines has also previewed Inkling-Small, a 276-billion-parameter model with 12 billion active parameters. The company says the smaller version matches or exceeds the flagship on several tests, but its full weights have not yet been released. For many developers, that model may be more practical than the flagship because Inkling’s hardware requirements place it beyond ordinary workstations.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab announcement
ESP32 Basic Starter Ai Chatbot Kit Development Board USB-C Dual Core Microcontroller Support AP/STA/AP+STA Compatible with Arduino IDE IoT for Beginners Engineers
【High Performance】The ESP32 module is equipped with a dual-core CPU and features a Type-C USB interface, with 44…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and License Need Checks
The release does not disclose Inkling’s training dataset or full training pipeline, meaning the model is open-weight but not fully open-source. Its benchmark scores are also vendor-published results, and several comparisons have not been reproduced by independent evaluators.
A separate Model Acceptable Use Policy has been reported as applying to the parameters and modified versions, with restrictions involving surveillance, deception and automated decisions affecting rights. That policy was not verified in the supplied reporting. Users will need to examine the current model card and legal terms because Apache 2.0 may not describe every applicable condition.
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.
Independent Tests and Smaller Weights
Researchers and prospective adopters will next test Inkling’s benchmark claims, reasoning-effort controls and multimodal performance on independent workloads. Attention will also turn to the promised Inkling-Small weight release, updated licensing information and evidence showing whether the model’s efficiency claims hold under sustained deployment.
AI model licensing tools
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 full BF16 and NVFP4 weights for its first foundation model, Inkling, through Hugging Face under Apache 2.0.
Why does releasing weights before an API matter?
It allows capable organizations to host, modify and fine-tune the model without depending entirely on a closed service. That can provide greater control over data and deployment.
Can Inkling run on a normal workstation?
No. The BF16 version reportedly needs at least 2 terabytes of aggregate VRAM, while NVFP4 still requires about 600 gigabytes. Those requirements limit direct deployment to well-funded operators with specialized hardware.
Is Inkling the strongest available AI model?
No. Thinking Machines itself says Inkling is not the strongest open or closed model. Its reported strengths include mathematics, adversarial robustness and audio, while it trails rivals on several coding and agent benchmarks.
Are Inkling’s performance results confirmed?
The reported scores come from the developer and associated evaluations. They offer an initial comparison, but independent replication remains pending, and some tests used a prerelease checkpoint.
Source: Thorsten Meyer AI
