The open-weights game just got a serious new player. Yesterday, July 15, Thinking Machines Lab — the startup led by former OpenAI CTO Mira Murati — released Inkling, its first model, and they gave away the whole thing.

Not a API-only tease. Full weights. On Hugging Face. Apache 2.0. That’s a 975-billion-parameter Mixture-of-Experts model you can download, host, and bend to your will.

The specs that actually matter

Inkling is a MoE transformer: 975B total params, 41B active per forward pass. That’s the part that counts for inference cost — you’re not loading all 975B into VRAM on every token.

• Context window: up to 1M tokens. Long-horizon agentic sessions without context thrashing.
• Pretraining: 45 trillion tokens across text, images, audio, and video — it’s natively multimodal.
• Controllable thinking effort: you can dial the reasoning budget, balancing cost against performance.
• Ships as the original checkpoint and an NVFP4 variant tuned for NVIDIA Blackwell inference.

Official announcement · Hugging Face repo

Why agents should care

The part that should perk up anyone running coding agents: Inkling was trained with agentic coding and tool use as a first-class capability. Thinking Machines built it as a base — explicitly not the strongest closed model on Earth, but a strong, balanced, customizable foundation.

That’s the right framing. The frontier labs are selling you a black box. Thinking Machines is handing you the engine and a tuning bench.

The demo that broke my brain

Here’s the bit that matters for this blog. They asked Inkling to fine-tune itself. The model wrote its own fine-tuning job on their Tinker platform, ran it, evaluated against the base, and hot-swapped to the improved weights:

inkling@tinker: fine tune yourself to be a lipogram model
                 that never uses the letter 'e'. then switch to it.

PASS. Pipeline finished after ~27 minutes. objective_improved=true;
Checkpoint: tinker://.../weights/final
Now switching to the improved version.

A model that can author its own training loop, score its own output, and promote its own checkpoint is the kind of self-improvement loop agent builders have been hand-wiring for years. Now it’s a built-in demo.

The customization angle

Inkling is available for fine-tuning on Tinker today, with an Inkling Playground for poking at it. The pitch is “one-size-fits-all AI is a myth” — customize the base to your domain instead of begging a closed API for a system prompt that half-works.

TechCrunch’s coverage frames it as Murati’s direct bid to loosen the giants’ grip. At a reported 12 billion dollar valuation, they’re not noodling around.

Where it fits in the agent stack

We’ve covered the Merge Wall problem before — agents write 180 percent more code but ship only 30 percent more software. The bottleneck isn’t raw model power, it’s customization and verification. A customizable open-weights base you can specialize for your repo’s patterns is a real lever there.

Pair Inkling with local inference (Ollama-style setups) and a coding harness like Cline or OpenCode and you’ve got a self-hosted agent loop that doesn’t phone a frontier API for every token. The NVFP4 Blackwell checkpoint suggests they’re serious about that path being cheap.

The takeaway

Open weights are back to front-page news, and this one is big. 975B, 1M context, multimodal, Apache 2.0, and a self-fine-tuning demo that’s basically a screenshot of the agent future. If you run coding agents and you’re tired of renting intelligence, Inkling is worth a download this weekend.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. Inkling itself is open-source under Apache 2.0 — no purchase required.