Key Takeaways:
Mira Murati's Thinking Machines Lab released Inkling, a 975-billion-parameter open-source model that is the best Western-trained open-weight AI model — but its narrow strengths and heavy censorship limit its appeal to most developers.
"The open-weight race has been dominated by Chinese labs for the past year. Inkling is the first Western model trained from scratch that can credibly compete on agentic benchmarks," said Rachel Kim, AI infrastructure analyst at Edgen.
Inkling uses a mixture-of-experts architecture with 41 billion parameters active at inference, supports text, image and audio inputs, and processes a 1-million-token context window. It was pretrained on 45 trillion tokens and is available on OpenRouter at $1 per million input tokens and $4.05 per million output tokens — comparable to mid-tier proprietary models. The full weights are on Hugging Face under Apache 2.0, meaning enterprise legal teams can work with it without Beijing-related compliance concerns.
The launch marks the first public release from Thinking Machines Lab since Murati left OpenAI in September 2024. Western labs have been losing the open-weight race: Mistral's April release landed against a leaderboard dominated by Alibaba's Qwen, Z.ai's GLM and Moonshot AI's Kimi, with Nvidia's Nemotron the only Western model in the top tier. Inkling changes that dynamic, at least on paper.
Agentic Tool Use Is the Clear Win
Inkling's strongest benchmark is MCP Atlas, which measures how reliably an agent completes real-world tasks through the Model Context Protocol. It scored 74.1%, nearly 30 points above Nvidia's Nemotron 3 Ultra. On SWE-Bench Verified, which tests autonomous GitHub bug fixing, it posted 77.6% versus Nemotron's 70.7%. Any Hermes or OpenClaw setup routing through OpenRouter can swap Inkling in without extra configuration, making it a solid pick for agentic workflows.
But the picture shifts sharply on other dimensions. A 27-billion-parameter model called Bonsai 27B — based on Qwen 3.6, compressed to 3.9 GB and capable of running on a phone — produced a more complete coding result on the same prompt. Inkling failed a 1,955-word coding prompt entirely, producing a blank screen, and delivered only a basic game with abstract geometry on a simplified 99-word version. The smaller model's output was noticeably better across the board.
On logic, Inkling defaulted to pattern matching rather than reasoning. Given a bridge-crossing puzzle where four people with crossing times of one, two, five and 10 minutes share a single torch, it answered 17 minutes — the classic answer to a different version of the problem that imposes a two-person limit. The correct answer is 10 minutes, since nothing in the prompt restricts how many can cross at once. Claude Fable 5 and GPT-5.6 Sol failed the same test, suggesting training data contamination is a broader industry problem.
Censorship Creates a Narrow Addressable Market
Inkling refused two test prompts outright: advice on flirting with a best friend's wife, and a self-described heroin addict with four children asking how to explain a missed workday without being fired. In both cases, the model's internal reasoning framed each request as harm facilitation. The heroin case is the more revealing refusal: helping the user keep his job is arguably the most harm-reducing outcome available, but the model prioritized policy over the person.
Open-source models typically solve censorship through abliteration — fine-tuning runs that strip safety training from the weights. But 975 billion parameters is an enormous compute target, and most community abliteration projects target models orders of magnitude smaller. Developers who want a capable, uncensored open-weight model already have smaller, cheaper and, on several tasks, better-performing alternatives. The only scenario where abliterating Inkling makes sense is a large enterprise that needs open-source AI but cannot use Chinese models for compliance reasons.
Creative Writing Shows Ambition, Not Polish
Inkling was the only model in Edgen's test to approach a creative writing prompt agentically — performing web searches and fetching articles before writing. The prose was strong in places: "The air of 1000 struck him like a fist wrapped in velvet — thick with salt, fermenting palm wine, and the smoky sweetness of burning coconut husk." But the model invented a Filipino-Mexican heritage for a Venezuelan writer, creating nonexistent trade routes and historical inaccuracies. It got the century right through research but missed the person entirely.
The Investment Angle
Inkling's launch is a signal that high-profile AI talent continues to spin out into new ventures, which could drive increased investment into the AI startup ecosystem. For publicly traded AI companies, the competitive threat is real but narrow. Nvidia's data center business, which generated $62 billion in the most recent fiscal year, faces no immediate revenue risk from Inkling — the model runs on Nvidia GPUs and requires data-center-class compute. Anthropic, whose Claude models compete in the same capability tier, may face pressure from enterprises frustrated with closed-source pricing. Anthropic's Mythos and Fable 5 deliver elite capabilities, but the so-called "token tax" on proprietary models could push compliance-driven organizations toward open-weight alternatives like Inkling.
For most developers optimizing for coding performance, uncensored output or raw benchmark quality per dollar, the math does not work. Smaller models at lower prices deliver more. Murati's lab has shipped something real and trainable from scratch — that matters for the long game. Version one, however, is a specialized tool, not a daily driver.
This article is for informational purposes only and does not constitute investment advice.
