TL;DR
Thinking Machines Lab released the full weights for its first foundation model, Inkling, before offering a closed API. The Apache 2.0 release gives developers broad control, but the model requires costly hardware, trails rivals on several benchmarks and may carry separate use restrictions.
Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, released the full weights for Inkling, its first foundation model, on July 15 under an Apache 2.0 license. Publishing the weights before any closed API gives organizations direct control over deployment and modification, although the lab acknowledged that Inkling is not the strongest model available.
The release includes BF16 and NVFP4 checkpoints on Hugging Face, with launch-day support for Transformers, vLLM, SGLang and llama.cpp, according to the model materials summarized by Thorsten Meyer AI. Apache 2.0 generally permits downloading, modifying and commercially using the weights, subject to the license terms and any other applicable conditions.
Inkling is described as a mixture-of-experts transformer with 975 billion total parameters and 41 billion active parameters. Thinking Machines says the model has a one-million-token context window and was pretrained on 45 trillion tokens spanning text, images, audio and video. It accepts text, images and audio while producing text.
The company also introduced an effort control ranging from 0.2 to 0.99, allowing operators to trade reasoning work against latency and token use. On Terminal-Bench 2.1, the model reportedly matched Nemotron 3 Ultra while using about one-third as many tokens, but that result comes from launch reporting and has not been independently replicated.
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 Weights Shift Model Control
Releasing weights at launch gives companies a path to run Inkling on their own infrastructure, modify it for specialized work and reduce dependence on a provider-controlled endpoint. That can matter for organizations focused on data residency, service continuity and model customization.
The release also tests whether buyers will value ownership and deployment flexibility over first place on benchmark tables. Thinking Machines openly conceded that Inkling does not lead the full market, framing its license, tooling and adjustable compute as central parts of the offering.
GPU Kernel Engineering for LLM Inference: CUDA, Triton, and Flash Attention Optimization for High-Throughput AI Production Systems (AI Infrastructure, Hardware & Compiler Engineering Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Murati’s Lab Chooses Openness
Thinking Machines Lab is 17 months old and includes researchers who previously worked on ChatGPT, according to the supplied report. Inkling is its first foundation-model release, making the decision to publish weights immediately an early indication of the company’s commercial and technical strategy.
The model’s reported results are mixed. Vendor figures place it at 97.1% on AIME 2026 and 87.2% on GPQA Diamond, while it trails named rivals on SWE-bench Pro, Terminal-Bench 2.1 and Humanity’s Last Exam. Some scores were produced with a pre-release checkpoint, limiting direct comparisons.
Rosewill 4U Server Chassis Case|Supports up to 4 GPUs|8 Hot-Swap 3.5"/2.5" SATA/SAS up to 12Gbps|E-ATX Compatible|3x 12038 Hot-Swap Fans,2 Rear 8038 Fans|USB 3.2 Type-C|With Rail Kit-RSV-AI01
AI-Optimized: Designed to support up to 4 GPUs, it is perfect for handling intensive AI and machine learning…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Policy and Performance Need Verification
It is not yet clear whether a reported Model Acceptable Use Policy places additional restrictions on the weights or modified versions. The source says the policy may bar surveillance, deception and fully automated decisions affecting rights, but it labels that account unverified. Developers will need to read the current model card and repository terms before deployment.
The published benchmark results also await independent replication. Training data and the full training pipeline were not released, meaning Inkling is open-weight rather than fully open-source. The available material does not provide enough information to independently examine dataset composition, filtering or all safety-training decisions.
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 customers are expected to test Inkling on real workloads, compare its cost curve with GLM-5.2, Kimi K2.6 and closed services, and examine whether its adjustable reasoning setting produces consistent savings. Legal teams will also need to resolve the reported use-policy question.
Attention will then turn to Inkling-Small, a preview model with 276 billion total parameters and 12 billion active. Thinking Machines says its full weights will follow after testing. Even that version may require substantial hardware, but its lower active-parameter count could make it more practical for a wider group of operators.
NOVATECH Apex AI Workstation & Gaming PC – AMD Ryzen 9 9950X3D, Machine Learning, Data Science, 3D Rendering, Video Editing, Simulation (RTX 5090 | 96GB RAM | 5TB)
[High-Performance AI & Machine Learning] The AMD Ryzen 9 9950X3D paired with the RTX 5090 (32GB VRAM) makes…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are model weights?
Model weights are the learned numerical values that shape how an AI system processes inputs and generates outputs. Access to them allows qualified operators to host, inspect and modify a model without relying solely on the developer’s API.
Is Inkling fully open-source?
No. The weights are available under Apache 2.0, but the training data and full training pipeline have not been published. That makes Inkling an open-weight release rather than a fully reproducible open-source system.
Can Inkling run on a personal workstation?
Not in its standard forms. The source estimates that BF16 requires at least two terabytes of aggregate VRAM, while NVFP4 still needs about 600 gigabytes. Heavily compressed community versions may lower the requirement while reducing accuracy.
Does Inkling lead current AI benchmarks?
No. It reports strong mathematics, audio and calibration results, but trails other models on several coding, agent and reasoning tests. The figures are vendor-published and awaiting independent replication.
Why release weights before an API?
The order gives developers direct possession of the model from the start and signals that self-hosting and customization are part of the core product strategy. Whether that approach gains broad adoption will depend on hardware cost, licensing clarity and real-world performance.
Source: Thorsten Meyer AI
