TL;DR
Thinking Machines Lab released its first foundation model, Inkling, with full weights under Apache 2.0 and immediate support from major inference frameworks. The lab concedes that Inkling is not the strongest available model, while emphasizing ownership, deployment flexibility and adjustable reasoning costs.
Thinking Machines Lab, the 17-month-old company founded by former OpenAI chief technology officer Mira Murati, released its first foundation model, Inkling, on July 15 with full weights available immediately on Hugging Face under an Apache 2.0 license. The open-first release gives organizations a path to modify and operate the model themselves, although its hardware demands place the flagship beyond most individual developers and smaller teams.
Inkling is a mixture-of-experts model with 975 billion total parameters and 41 billion active parameters, according to specifications published by Thinking Machines. The company says it supports a 1-million-token context window and was pretrained on 45 trillion tokens spanning text, images, audio and video.
The model accepts text, images and audio and produces text. Thinking Machines released BF16 and NVFP4 checkpoints with day-zero support in Transformers, vLLM, SGLang and llama.cpp, among other tools. The weights and model card carry an Apache 2.0 license, which generally permits modification and commercial use.
The company did not claim overall market leadership. Its announcement said Inkling is “not the strongest model available today”, whether compared with open or closed systems. Vendor-published results place it ahead on some mathematics, audio and adversarial tests but behind models including GLM-5.2 and Fable 5 on several coding and agent-oriented evaluations. Those results have not yet received broad independent replication, and some reportedly used a prerelease 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.
Open Distribution Changes Model Control
Releasing the weights first makes customer ownership, rather than API access, a central part of the product. Organizations can inspect, modify, fine-tune and host Inkling within their own infrastructure, reducing dependence on a provider that can change prices, access rules or model versions.
Inkling also includes an adjustable reasoning-effort setting from 0.2 to 0.99. Thinking Machines says this lets operators trade reasoning tokens against latency and cost. The company reports that Inkling matched Nemotron 3 Ultra on Terminal-Bench 2.1 while using about one-third as many tokens, but that comparison remains a vendor claim pending outside testing.
The FPGA Programming Handbook: An essential guide to FPGA design for transforming ideas into hardware using SystemVerilog and VHDL
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Murati’s Lab Chooses Open First
Thinking Machines was founded by Murati and former OpenAI employees, including people who worked on ChatGPT. Many frontier-model developers distribute their strongest systems through controlled APIs, while open weights arrive later, apply only to smaller models or are released under more restrictive terms. Inkling reverses that sequence by putting downloadable checkpoints first.
The release also enters a field where Chinese developers have supplied several leading open-weight models. Inkling is positioned as a US-developed alternative, yet the source material reports that its post-training used synthetic data from Kimi K2.5. Thinking Machines has not published Inkling’s training dataset or full pipeline, meaning the release is open-weight rather than fully open-source.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab
High-Performance AI Systems Engineering: Techniques for Faster Model Training, Efficient GPU Workloads, Distributed Computing, and Reliable AI Deployment across Modern Infrastructure
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License Limits and Benchmarks Need Checks
It is not yet clear whether a separate Model Acceptable Use Policy restricts the parameters and modified versions beyond the Apache 2.0 license. The source material reports possible bans covering surveillance, deception and fully automated decisions affecting rights, but says the policy was not independently verified. Prospective users will need to examine the current repository documents before deployment.
Independent evaluators have also not fully tested the company’s benchmark, efficiency and multimodal claims. Inkling’s real-world reliability, fine-tuning behavior and operating costs across different hardware configurations remain uncertain.
Building MCP Servers for AI Agents: Scalable Architecture Patterns, Security Design, and Production-Ready AI Infrastructure for Large Language Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Smaller Weights Follow
Researchers and prospective customers will now test Inkling’s released checkpoints against GLM-5.2, Kimi K2.6 and closed systems on production workloads. Thinking Machines is also testing Inkling-Small, a 276-billion-parameter model with 12 billion active parameters, and says its full weights will follow after testing is complete.
NVD RTX PRO 6000 Blackwell Professional Workstation Edition Graphics Card for AI, Design, Simulation, Engineering – 96GB DDR7 ECC Memory – 4th Gen RT/5th Gen Tensor Core GPU – OEM Packaging
[NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Thinking Machines release?
The company released Inkling, its first foundation model, including BF16 and NVFP4 weights hosted on Hugging Face.
Is Inkling fully open-source?
No. Its model weights are available under Apache 2.0, but the training data and complete training pipeline have not been published.
Can Inkling run on a personal workstation?
Not in its standard released forms. The source estimates that BF16 needs at least 2 terabytes of aggregate VRAM, while NVFP4 still requires about 600 gigabytes. Quantized versions may reduce the requirement, with possible quality losses.
Is Inkling the best-performing open model?
Thinking Machines says it is not the strongest model overall. Vendor results show competitive performance on selected tests, but Inkling trails some rivals on coding, agent and general reasoning benchmarks.
When will Inkling-Small be available?
Thinking Machines has shown a preview and says the weights will be released after testing. The company has not provided a firm release date.
Source: Thorsten Meyer AI
