TL;DR
Thinking Machines Lab released the full weights for Inkling, its first foundation model, on July 15 under the Apache 2.0 license before offering a closed API. The release gives organizations more control, but large hardware demands, a reported separate use policy and unverified benchmark claims limit what can be concluded.
Thinking Machines Lab, founded by former OpenAI technology chief Mira Murati, released the full weights for its first foundation model, Inkling, on July 15 under Apache 2.0 before offering a closed API. The release matters because it gives developers direct access from launch, while the lab also acknowledged that Inkling is not the strongest available model.
Inkling is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active parameters. According to the release materials summarized by Thorsten Meyer AI, it has a one-million-token context window and was pretrained on 45 trillion tokens spanning text, images, audio and video.
The lab published BF16 and NVFP4 checkpoints on Hugging Face, with launch-day support for Transformers, vLLM, SGLang and llama.cpp. Apache 2.0 permits downloading, modifying and commercially using the weights. The release does not include the training dataset or full training pipeline, meaning Inkling is an open-weight model rather than a fully open-source system.
Vendor-reported results place Inkling at 97.1% on AIME 2026, 87.2% on GPQA Diamond and 91.4% on VoiceBench. It trails cited competitors on several software and agent benchmarks, including SWE-bench Pro and Terminal-Bench 2.1. These figures have not been independently replicated, and some reportedly came from a pre-release checkpoint.
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.
Ownership Comes With Infrastructure Costs
Releasing the weights first allows organizations to host Inkling themselves, inspect its behavior, modify it and avoid dependence on a single API provider. That could appeal to governments, regulated industries and research institutions seeking greater control over model access, data handling and deployment.
Access to the files does not make the flagship broadly accessible. Thorsten Meyer AI estimates that BF16 deployment requires at least 2 terabytes of aggregate VRAM, while NVFP4 still needs about 600 gigabytes. Those requirements put practical deployment beyond most workstations and make infrastructure cost part of any ownership calculation.
Inkling also offers a 0.2-to-0.99 reasoning-effort setting intended to trade computation for speed and cost. The source reports that the model matched Nemotron 3 Ultra on Terminal-Bench 2.1 using roughly one-third of the tokens, but that comparison awaits independent testing.
HP 15.6" Laptop – Complete Productivity Solution, Windows 11, Microsoft Office, Intel 4 Core N100, 16GB RAM, 640GB Storage (128GB UFS + 512GB SD Card), Copilot AI, Lightweight – Silver
- Memory: 16GB high-speed RAM for multitasking
- Storage: 640GB total storage including UFS and SD card
- Display: 15.6-inch HD screen for clear visuals
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Murati’s Lab Chooses Openness First
Thinking Machines Lab was founded about 17 months before the release by Murati and employs former OpenAI personnel who worked on ChatGPT, according to Thorsten Meyer AI. Rather than beginning with paid API access, the company made downloadable weights the first public form of its model.
The release arrives as American developers seek alternatives to Chinese open-weight models including GLM-5.2 and Kimi K2.6. The supplied benchmark comparisons indicate that those systems retain leads in some reasoning, agent and multimodal tasks. Inkling’s post-training also reportedly used synthetic data from Kimi K2.5.
The lab separately previewed Inkling-Small, a 276-billion-parameter model with 12 billion active parameters. Thinking Machines Lab said its weights will be published after testing, potentially giving more organizations a less demanding deployment option.
“Open weights used to be a consolation prize. Inkling is a strategic open release.”
— Thorsten Meyer AI
Licensing and Performance Need Verification
It remains unclear whether a separate Model Acceptable Use Policy places additional restrictions on the weights or modified versions. The supplied source says such a policy has been reported but was not independently verified. Developers in surveillance, geospatial analysis, public safety or automated rights decisions would need to check the current model card and legal terms.
The model’s real-world performance is also unsettled. Most cited benchmarks were published by the vendor or reported through third parties, and no broad body of independent replication was available at publication. The effect of hardware configuration, quantization and reasoning-effort settings on cost and accuracy is also unknown.
Independent Tests and Smaller Weights Follow
Researchers and prospective users will now test Inkling on their own workloads, compare it with GLM-5.2, Kimi K2.6 and other open models, and examine whether its cost-versus-performance curve supports production deployment.
Attention will also turn to the final legal terms and the promised release of Inkling-Small’s full weights. Those developments will show whether the lab’s open-first approach reaches beyond well-funded operators with large accelerator fleets.
Key Questions
What did Thinking Machines Lab release?
The company released Inkling’s full BF16 and NVFP4 weights on Hugging Face under Apache 2.0, alongside support for several widely used inference tools.
Is Inkling fully open source?
No. The model weights are open, but the training dataset and full training pipeline were not published. A separate use policy has also been reported but was not verified in the supplied material.
Can Inkling run on a consumer workstation?
The flagship is unlikely to run at full capability on ordinary hardware. Reported requirements are at least 2 terabytes of VRAM for BF16 or about 600 gigabytes for NVFP4.
Is Inkling the highest-performing open model?
No. Thinking Machines Lab said Inkling is not the strongest model available. Vendor results show strengths in mathematics, audio and calibration, but competitors lead several coding and agent benchmarks.
Why does releasing weights first matter?
It lets organizations download, modify and self-host the model without waiting for a closed service. That offers more deployment control, although hardware and legal constraints may still limit adoption.
Source: Thorsten Meyer AI
