Thinking Machines' Inkling bets on customisable open weights, not frontier scores
· TexaHub · LLMs & Foundation Models, Model Comparisons, Artificial Intelligence
Mira Murati's lab has released a 975-billion-parameter open-weight model under Apache 2.0, pitching it as a base for fine-tuning while US debate over open models intensifies.
The release
Thinking Machines Lab released Inkling on 15 July, its first model trained from scratch with full weights available for download [1]. The company describes it as a mixture-of-experts transformer with 975 billion total parameters, of which 41 billion are active per token, a context window of up to 1 million tokens, and pretraining on 45 trillion tokens of text, images, audio and video [1]. Alongside it, the lab shared a preview of Inkling-Small, a lighter model with 12 billion active parameters [1].
The weights are published on Hugging Face under the Apache 2.0 licence, both as the original checkpoint and in NVFP4 format for Nvidia Blackwell systems [1][3]. The model card describes a general-purpose model that accepts text, image and audio input and produces text output [3]. Thinking Machines, founded by former OpenAI chief technology officer Mira Murati, has around 200 employees, according to TechCrunch [2].
How it compares
Unusually for a launch post, the company states plainly that Inkling "is not the strongest overall model available today, open or closed" [1]. Its own benchmark table supports that. On SWE-bench Verified, Inkling scored 77.6%, against 80.2% for Moonshot's Kimi K2.6, 80.0% for GLM 5.2 and 80.6% for DeepSeek V4 Pro among open models, and 82.2% for GPT 5.6 Sol and 95.0% for Claude Fable 5 among closed ones [1]. On the text-only version of Humanity's Last Exam it scored 29.7%, compared with 40.1% for GLM 5.2, 47.2% for GPT 5.6 Sol and 53.3% for Claude Fable 5 [1]. On Terminal Bench 2.1 the gap is similar: 63.8% for Inkling against 82.7% for GLM 5.2 and 89.5% for GPT 5.6 Sol [1]. On the GPQA Diamond science benchmark it scored 87.2%, against 94.1% for both GPT 5.6 Sol and Gemini 3.1 Pro [1]. None of this is far off the open-weight pack, but it is clearly short of the closed frontier.
There are areas where it leads. Inkling's IFBench instruction-following score of 79.8% is above the 72.7% listed for GPT 5.6 Sol and the 63.5% for Claude Fable 5, though slightly below Nvidia's Nemotron 3 Ultra at 81.4% [1]. The company also claims efficiency gains from a controllable thinking-effort setting, saying Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 using roughly a third of the tokens [1]. These are vendor-run numbers; Thinking Machines says it relied on Artificial Analysis scores for several evaluations to improve consistency, and that the Humanity's Last Exam figures reflect an earlier checkpoint [1].
The model card is candid about limitations, listing hallucination, occasional failures to follow instructions precisely and degraded performance in long multi-turn conversations [3].
The customisation pitch
The commercial logic sits in Tinker, the company's fine-tuning platform, where Inkling was available from launch day [1]. Thinking Machines argues that multimodal capability, efficient reasoning and Tinker access together make Inkling a good base for customisation [1]. It said the model is also served through Together AI, Fireworks, Modal, Databricks and Baseten, with inference support in vLLM, SGLang and llama.cpp [1]. TechCrunch reported a collaboration with Bridgewater Associates in which a customised model reached 84.7% on a financial reasoning task at roughly one-fourteenth the operating cost of proprietary models [2]; that is a single customer case reported second-hand, not an independent benchmark.
On training, the lab said Inkling's mixture-of-experts design largely follows DeepSeek-V3, with 256 routed and two shared experts per layer, and that it was trained on Nvidia GB300 NVL72 systems [1]. It reported scaling reinforcement learning to more than 30 million rollouts [1].
The bet matches a broader shift in how models are used. TechCrunch, citing Hugging Face, reported that Chinese open-weight models accounted for 41% of downloads on the platform in spring 2026, ahead of US models, and that open-weight models handled nearly a third of AI requests on Vercel in June [4]. A US-built, Apache-licensed model in this size class gives enterprises that want to own their weights an option that is not Chinese-made, which may matter more to some buyers than leaderboard position.
Safety and the policy backdrop
Thinking Machines said it trained Inkling to an internal safety specification and commissioned external testers covering CBRN, cyber and loss-of-control risks [1]. It reported a 78.0% score on the adversarial portion of the FORTRESS refusal benchmark, the highest among the open models it compared, and above 98% on StrongREJECT [1]. The company also acknowledged that it is still studying how fine-tuning on Tinker affects safety behaviour [1], which is the central question for any open-weight release, since safeguards can be trained away once weights are public.
The release landed in a heated week. Moonshot AI's Kimi K3, released days later, was described by Moonshot as showing frontier-level performance, though still behind Claude Fable 5 and GPT 5.6 Sol, and TechCrunch linked a roughly 1% Nasdaq drop that Friday to selling in chip stocks [5]. TechCrunch also reported that, according to Axios, the administration was considering banning K3 and other advanced Chinese models, while Politico reported that the Commerce Department might not pursue such restrictions [6].
For Thinking Machines, any broad restriction on open weights would cut against its model; a restriction aimed at Chinese models could do the opposite and widen demand for US alternatives. Which way Washington goes is unresolved. In the meantime, Inkling's practical test is whether developers fine-tune it in numbers, which download counts and Tinker usage will show over the coming months.
Sources
[1] Inkling: Our Open-Weights Model — Thinking Machines Lab, 15 Jul 2026
[2] Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling — TechCrunch, 15 Jul 2026
[3] thinkingmachines/Inkling model card — Hugging Face, undated
[4] The real AI race may no longer be at the frontier — TechCrunch, 14 Jul 2026
[5] Kimi: Threat or menace? — TechCrunch, 18 Jul 2026
[6] OpenAI is scared of open-weight models. Should the US be? — TechCrunch, 20 Jul 2026