TL;DR
Thinking Machines Lab released the full weights for its first foundation model, Inkling, before offering a closed API. The model introduces adjustable reasoning effort but trails rivals on several vendor-reported benchmarks, while its licensing and hardware demands require closer review.
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 an Apache 2.0 license. By publishing downloadable weights before launching a closed API, the 17-month-old company is testing whether model ownership and deployment control can matter more to customers than leading every benchmark.
Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters, according to materials cited in the supplied reporting. It accepts text, images and audio, produces text, and supports a 1-million-token context window. The lab says it pretrained the model on 45 trillion tokens spanning text, images, audio and video.
The BF16 and NVFP4 checkpoints appeared on Hugging Face with immediate support for Transformers, vLLM, SGLang and llama.cpp. Apache 2.0 generally permits downloading, modifying and commercially deploying the weights. The company did not publish its training data or full training pipeline, making Inkling an open-weight release rather than a fully open-source model.
A configurable thinking-effort setting, ranging from 0.2 to 0.99, lets operators trade more reasoning tokens for higher cost and latency. Thinking Machines reportedly found that Inkling matched Nemotron 3 Ultra on Terminal-Bench 2.1 while using about one-third as many tokens, but that comparison has not received independent confirmation.
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 Before Benchmark Rank
The release gives companies a Western-developed foundation model that can be hosted, modified and fine-tuned without depending solely on a vendor-controlled API. That may appeal to buyers focused on data control, service continuity and custom deployment.
The adjustable effort setting also shifts attention from a single leaderboard score to a cost-performance curve. For high-volume services, the ability to reduce reasoning tokens on routine requests may matter more than maximum performance. It does not show that Inkling thinks like a person or possesses awareness; it shows that inference computation can be controlled and measured against latency, price and task accuracy.
Samsung 43-Inch Class U8000H Series Crystal UHD, 4K, Smart TV, 2026 Model
- Crystal Processor: Enhances colors and sharpness in 4K
- Free Streaming Content: Access 2,700+ free channels and shows
- Motion Xcelerator: Smooth motion for gaming and sports
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Inside Inkling’s Open Release
Thinking Machines Lab was created by Murati and employs several people who previously worked on ChatGPT. Its first model enters an open-weight market where Chinese systems including GLM-5.2 and Kimi K2.6 already perform strongly on reasoning, agentic and multimodal tests.
Vendor-published results place Inkling at 97.1% on AIME 2026 and 87.2% on GPQA Diamond. The supplied analysis also reports lower results than selected rivals on Humanity’s Last Exam, SWE-bench Pro and Terminal-Bench 2.1. Thinking Machines has acknowledged that Inkling is not the strongest available open or closed model.
“Inkling is not the strongest model available today, closed or open.”
— Thinking Machines Lab, as quoted in the supplied source
Designing Machine Learning Systems: An Iterative Process for Production-Ready Applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and License Need Verification
Most cited benchmark figures are vendor-published, some involve a prerelease checkpoint, and independent researchers have not yet replicated them. Claims about spontaneous chain-of-thought compression during reinforcement learning also require fuller technical evidence before broader conclusions can be drawn.
A separate Model Acceptable Use Policy has been reported as restricting surveillance, deception and fully automated decisions affecting rights, including for modified versions. The supplied source did not verify that policy. Users will need to examine the model card and repository because any extra conditions could affect whether Apache 2.0 describes the complete legal position.
AI reasoning effort adjustment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Smaller Model
Researchers and prospective users will now test Inkling on real workloads, inspect its license terms and compare its efficiency with GLM-5.2, Kimi K2.6 and closed models. Attention will also turn to Inkling-Small, a 276-billion-parameter model with 12 billion active parameters whose full weights are expected after testing. That smaller release may be more practical because the flagship reportedly needs at least 600 GB of memory in NVFP4 or about 2 TB for BF16 deployment.
The n8n Blueprint : Build AI Agents, Automate Your Business, and Turn No-Code Skills Into Income (Made Simple: Beginner's Guides to the Apps and Software Running Modern Work and Life)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What does a weights-first release mean?
It means downloadable model parameters arrived before a closed API. Qualified users can host and modify Inkling instead of sending every request through Thinking Machines Lab.
Is Inkling fully open source?
No. The weights are available under Apache 2.0, but the training dataset and complete pipeline were not published. Any separate use policy also needs verification.
Is Inkling the strongest AI model?
No. Thinking Machines says it is not the strongest model, and supplied benchmark results show it trailing some competitors on coding, reasoning and agentic tasks.
Can Inkling run on a personal workstation?
Not in its standard released forms. The reported requirements are about 600 GB of memory for NVFP4 and at least 2 TB for BF16. Smaller quantized versions may reduce that burden, with performance tradeoffs.
Source: Thorsten Meyer AI
