Mira Murati's first card
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, pulled the wraps off its first foundation model on July 17 — Inkling.
975 billion parameters, Mixture-of-Experts, 41B active, full open weights. Not a slide deck — an actual downloadable model.
The interesting part: Inkling isn't chasing the "biggest and strongest" leaderboard. It's pitching balance and customizability.
Native audio is the standout
Inkling is currently the largest open-weights model that natively supports audio. Pretrained on 45 trillion tokens spanning text, images, audio, and video.
Why does that matter? Because prior open models handling audio were either small or had audio bolted on after the fact. Native means audio isn't "translated" in — it's wired into the model's brain. Lower latency, sharper understanding.
- VoiceBench: 91.4%
- MMAU: 77.2%
Good news for anyone building voice agents, podcast transcription, or real-time translation.
Controllable thinking depth
One design choice is clever: thinking effort is tunable.
On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra at roughly one-third the tokens.
In other words: token savings. For teams running heavy inference, tokens are money — cutting two-thirds off is a real cost edge.
Safety scores are unexpectedly high
Open-weights models get grilled hardest on safety. Inkling posted 78.0% on the FORTRESS adversarial benchmark — the highest among open-weights models. StrongREJECT sits at 98.6%.
Not "absolutely safe," but Thinking Machines Lab did its homework on the "dare to open" front. Open weights means anyone can self-host, so the safety bar can't be low.
A few more hard numbers
- SWE-bench Verified: 77.6%
- GPQA Diamond: 87.2%
- AIME 2026: 97.1%
Coding, graduate-level Q&A, math competitions — none of them weak. But Murati's team keeps hammering one point: they're not chasing top of the leaderboard, they're after balance.
There's a small one too
Dropped alongside is Inkling-Small, 276B parameters (12B active), performing close to its bigger sibling on reasoning and agentic tasks — suited for cost-sensitive deployments.
Fine-tuning runs on Tinker; APIs are live on TogetherAI, Fireworks, Modal, Databricks, and Baseten. Open weights means you can run it locally and modify it yourself.
How to read it
Mira Murati's post-OpenAI startup has been watched closely for a long time. The first card plays out like this — not chasing #1 on the leaderboard, chasing balance and customizability. Pretty on-brand for her.
The open-weights field is getting crowded: Moonshot's K3 goes for scale, Meta's Llama goes for ecosystem, DeepSeek goes for price-performance, and now Inkling goes for native multimodal plus controllable cost. The strategies are diverging — good news for developers.
One caveat though: open weights isn't a free lunch. Self-hosting eats GPU, ops, and a security team's attention. The model is free; the cost of keeping it running hasn't dropped a cent.