TL;DR
Thinking Machines Lab released its first foundation model, Inkling, with full weights under Apache 2.0 before offering a closed API. The release signals a weights-first strategy, although its hardware demands, vendor benchmarks and reported use restrictions require further scrutiny.
Thinking Machines Lab, the AI company founded by former OpenAI technology chief Mira Murati, released its first foundation model, Inkling, with downloadable weights under the Apache 2.0 license before introducing a closed API. The July 15 release matters less as a claim to benchmark leadership than as a test of whether a major American lab can make model ownership the starting point rather than a later concession.
According to the laboratory’s release materials summarized by Thorsten Meyer AI, Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters. It has a 1-million-token context window and was pretrained on a reported 45 trillion tokens spanning text, images, audio and video. The model accepts text, image and audio inputs and produces text.
The release includes BF16 and NVFP4 checkpoints on Hugging Face, with initial support for transformers, vLLM, SGLang and llama.cpp. Users can download, modify and commercially deploy the weights under Apache 2.0. The training data and full training pipeline were not published, however, so the release is more accurately described as open-weight rather than fully open-source.
Thinking Machines Lab acknowledged that Inkling is not the strongest available model, whether compared with proprietary or open alternatives. Vendor-published results place it ahead on selected mathematics, scientific reasoning, tool-use, audio and adversarial tests, while showing weaker results on several coding and agent benchmarks. Those figures, including some produced with a pre-release checkpoint, have not yet received broad 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.
Ownership Becomes the Launch Strategy
The release reverses a familiar industry sequence in which laboratories sell access through a controlled cloud API and release weights later, if at all. By publishing Inkling’s weights immediately, Thinking Machines Lab gives qualified operators the ability to host, modify and fine-tune the model without depending on continuing API access or accepting future service changes.
That approach could appeal to governments, research institutions and companies seeking greater control over sensitive data or protection against a provider withdrawing access. Yet ownership remains practical only for well-funded operators: Thorsten Meyer AI estimates that BF16 deployment needs at least 2 terabytes of aggregate VRAM, while NVFP4 still requires roughly 600 gigabytes. Inkling is open to download, but not readily runnable on ordinary workstations.
The model also introduces a reported 0.2-to-0.99 reasoning-effort setting that lets operators trade output quality against token use, latency and cost. The laboratory reportedly found that Inkling matched Nemotron 3 Ultra on Terminal-Bench 2.1 while using about one-third as many tokens. That is a vendor comparison and needs outside testing, but controllable compute may matter more to high-volume users than a single maximum benchmark score.
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
Upgraded to 16GB High-Speed Memory Boosted to 16GB RAM for improved multitasking — switch between apps, browser tabs,…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Western Open-Weight Bid
Thinking Machines Lab was founded about 17 months before the release and employs several people who previously helped build ChatGPT. Inkling enters an open-weight market where Chinese models, including GLM-5.2 and Kimi K2.6, have posted stronger results on some reasoning, coding and multimodal evaluations.
The laboratory has positioned Inkling partly as a Western alternative with censorship-resistance training. The release materials also indicate that post-training used synthetic data from Kimi K2.5, illustrating how closely connected the competing model ecosystems remain. A preview of Inkling-Small, with 276 billion total and 12 billion active parameters, accompanied the flagship announcement, but its full weights are due only after testing.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab, in its launch announcement
AI Robotic Arm Kit Hiwonder SO-ARM101 Embodied Imitation Learning Open Source 6-Axis Robot Arm 12 High-Torque Bus Servo Motors AI Vision Recognition (Standard Kit, Included 3D Printed Part, Assembled)
【End-to-End Imitation Learning】Hiwonder SO-ARM101 robot arm is an embodied intelligent hardware platform compatible with the Lerobot open-source framework….
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Licensing and Benchmarks Need Verification
Several parts of the release remain unresolved. The reported benchmark results are largely vendor-published, and some used a pre-release checkpoint, leaving Inkling’s relative performance uncertain until independent evaluators test the public weights under comparable conditions.
Thorsten Meyer AI also reported a separate Model Acceptable Use Policy covering parameters and modified versions, with restrictions involving surveillance, deception and automated decisions affecting rights. That policy was not independently verified in the supplied reporting. If it applies as described, users will need to determine how it interacts with Apache 2.0 before building commercial, public-safety or geospatial systems.
The laboratory has not disclosed the underlying training dataset or complete pipeline. It is also unclear how reliably the reasoning-effort control will reduce real-world serving costs, or whether the smaller Inkling model will preserve the flagship’s capabilities on independent tests.
Generative AI on AWS: Building Context-Aware Multimodal Reasoning Applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Smaller Weights
The next evidence will come from independent benchmark replication, deployment tests using the published checkpoints and legal review of any additional use policy. Prospective users will need to compare Inkling with GLM-5.2, Kimi K2.6 and proprietary APIs on their own workloads rather than relying on headline scores.
Attention will also shift to the release of Inkling-Small’s full weights. With 12 billion active parameters, the smaller model may be more relevant to local and lower-cost deployments. Its release schedule, hardware requirements and final license conditions remain unconfirmed.
LLM Systems Engineering: Training and Building Large Language Models – Engineering AI Models Through Fine-Tuning, Continued Pretraining, and From-Scratch Development (AI Engineering)
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 laboratory released Inkling, a 975-billion-parameter Mixture-of-Experts foundation model, along with downloadable BF16 and NVFP4 weights under Apache 2.0.
Is Inkling fully open-source?
Inkling is open-weight, allowing users to download and modify its parameters. The training data and full training pipeline have not been published, so the release does not provide every component needed to reproduce the model.
Can Inkling run on a personal workstation?
Not in its standard released forms. Reported requirements reach at least 2 terabytes of aggregate VRAM for BF16 and about 600 gigabytes for NVFP4, putting deployment beyond ordinary consumer hardware.
Is Inkling the best-performing open model?
No established evidence supports that conclusion. The laboratory says Inkling is not the strongest model available, and rival models outperform it on several published tests. Its reported strengths still require independent confirmation.
Why does the weights-first release matter?
Publishing the weights first gives capable users direct control over hosting and modification without depending entirely on a vendor API. The value is limited by hardware costs, incomplete training disclosure and possible additional use restrictions.
Source: Thorsten Meyer AI
