Thinking Machines Released Inkling. They Say It's Not the Best. That's the Strategy.

Mira Murati was OpenAI's CTO when GPT-4 shipped. She left in late 2024, spent 18 months building in stealth, and this week Thinking Machines Lab released its first in-house model: Inkling. Their own announcement says it is "not the strongest overall model available today, open or closed." That's not a caveat. It's the thesis.

What Inkling Actually Is

Inkling is a mixture-of-experts model: 975B total parameters, 41B active per token, trained from scratch on 45 trillion tokens of text, images, audio, and video. Native multimodal reasoning across all four modalities, not adapters bolted on later. Context window of 1M tokens. Full weights on Hugging Face under Apache 2.0, so you can download, modify, and deploy on-premise with no vendor relationship required.

The architecture: a 66-layer decoder-only transformer with a sparse MoE feed-forward backbone. Each MoE layer has 256 routed experts plus 2 shared experts that fire on every token, with 6 routed experts activating per forward pass. Sigmoid router with an auxiliary-loss-free load-balancing bias. For a 975B total parameter system, the per-query compute lands closer to a 41B dense model, which changes the economics significantly.

Inkling-Small also dropped: 276B total, 12B active. On instruction-following benchmarks like IFBench, Small scores 83.4% against the big model's 79.8%. That kind of inversion happens in MoE models more than people expect, and it's worth running your actual use case against both before assuming bigger is better.

The Benchmark Numbers

Numbers that matter here:

• 97.1% on AIME 2026 (math reasoning)
• 87.2% on GPQA Diamond (graduate-level science questions)
• 29.7% on HLE text-only, 46.0% with tool use
• 73.5% on MMMU Pro (multimodal understanding)
• 78.0% on FORTRESS Adversarial

The HLE number with tools is strong for an open-weight model. But comparing these against GPT-5.6 Sol or Sonnet 5 and treating the gap as a decision input misses the point. Thinking Machines is not trying to win that game, and they're saying so plainly.

The Actual Bet

The Tinker platform is where the strategy lives. You start with Inkling's weights, fine-tune on your domain data via Tinker's training API, and deploy something fitted to your specific problem. A legal firm doesn't need a model that wins HLE. They need one that handles privilege review, citation extraction, and clause classification without hallucinating case law, running on their own infrastructure so confidential filings never leave the building.

Apache 2.0 means your fine-tuned variant is yours. No usage fees, no terms of service clauses that restrict competitive applications, no vendor policy change that breaks your deployment six months from now. In regulated industries, the vendor relationship itself is a compliance risk. Open weights eliminate that.

NVIDIA backing Thinking Machines is also legible in this context. NVIDIA wants workloads running on GPU clusters. An enterprise that fine-tunes Inkling on its own hardware is a better NVIDIA customer than one paying OpenAI or Anthropic for API credits.

The Censorship Angle

Thinking Machines made a specific claim about what they call Inkling's "epistemics": calibration, instruction following, and resistance to censorship. On Cognition's Propaganda and Censorship Eval, Inkling showed strong non-compliance with prompts designed to make it refuse or distort factual outputs on sensitive topics.

I haven't run this in production. But the posture addresses a real friction with frontier closed models, which have grown more conservative as they've scaled. For medical second-opinion tooling, security research, investigative journalism, or legal analysis that touches contested facts, that conservatism creates gaps. Inkling is explicitly positioning to fill them.

What This Changes for Builders

The honest question is: when does fine-tuning Inkling beat calling a frontier API?

The calculation favors Inkling when you have a large, well-defined corpus, high query volume where per-token costs compound, data residency requirements, or a domain where a general model's caution is a liability. It probably doesn't make sense for low-volume general tasks, or when you need the best available reasoning without the infrastructure overhead of running fine-tuning pipelines.

Most teams right now are calling APIs because it's faster to start. That's fine. But as usage scales and the per-token bill grows, the calculus shifts. The tooling for self-hosted fine-tuning keeps improving, and models like Inkling keep narrowing the capability gap. You don't have to make this call today, but you should understand when the break-even point arrives.

Not a Finished Product

Thinking Machines said Inkling is a starting point. That framing is honest in a way most model launches aren't. They're not claiming to beat Sonnet 5 on a leaderboard. They're building customization infrastructure and betting that enterprise value will accrue to whoever owns the fine-tuned layer, not whoever runs the biggest pretraining cluster.

Whether that plays out depends on how many teams actually build fine-tuning workflows versus defaulting to API calls indefinitely. Right now it's mostly the latter. But the economics of API-first keep shifting, and Inkling is built for the moment when that shifts for your team.