Thinking Machines Inkling: Inside Mira Murati's First Open-Weights Model

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released its first in-house model: Inkling, an open-weights, multimodal Mixture-of-Experts system. The surprise isn't the size — it's the strategy. Inkling isn't built to win benchmarks; it's built to be downloaded, fine-tuned, and reshaped by the organizations that run it. This guide covers Inkling's specs, benchmarks, the "controllable thinking effort" idea, and what an open-weights base means if you're building AI agents.

What is Inkling?

Inkling is Thinking Machines Lab's first public model, released on July 15, 2026 — the lab's first real proof point after roughly 18 months of mostly private infrastructure work (TechCrunch).

The headline facts, from the lab's own release post:

• 975B total parameters, 41B active — a Mixture-of-Experts (MoE) transformer that activates only a fraction of its weights per token, which keeps a very large model faster and cheaper to run.
• 1M-token context window.
• Pretrained on 45 trillion tokens of text, images, audio, and video, reasoning natively across all four. For now, outputs are text: code, structured data, and styled artifacts.
• Open-weights — the full weights are on Hugging Face, both as the original checkpoint and an NVFP4 checkpoint for efficient inference on NVIDIA Blackwell.
• A preview of Inkling-Small, a 276B MoE (12B active) tuned for lower cost and latency.

The most striking line comes from the lab itself: Inkling is "not the strongest overall model available today, open or closed." That's the whole strategy, not a hedge.

The bet: fine-tune it yourself

Most frontier labs — OpenAI, Anthropic, Google — ship a general-purpose chatbot first, then layer agentic features on top. Inkling inverts that. Thinking Machines is marketing it less as a finished product and more as a base for customization: a broad, balanced foundation model that organizations adapt to their own workflows through Tinker, the lab's fine-tuning platform.

The argument, laid out in a company post that preceded the launch, is that a model trained centrally and then frozen underperforms one that each organization can shape around its own expertise. It's a view gaining outside support: Microsoft's Satya Nadella recently warned that firms relying on closed models effectively "pay twice" — once in subscription fees, and again by feeding proprietary knowledge into someone else's future model versions (TechCrunch).

The clearest proof point Thinking Machines offers is a project with hedge fund Bridgewater Associates: researchers fine-tuned an open model on Bridgewater's financial expertise, and it reportedly scored 84.7% on financial reasoning tests while costing roughly a fourteenth as much to run as top proprietary models. Worth flagging — those numbers come from the two companies' own evaluation, not an independent one.

The trade-off is ownership of risk. Because customers fine-tune Inkling themselves, they — not Thinking Machines — are responsible for keeping their customizations safe. And fine-tuning at this level takes real machine-learning talent.

Controllable thinking effort

The most agent-relevant feature is controllable thinking effort. Inkling lets you dial reasoning up or down — trading accuracy for speed and cost — and the setting can be adjusted from inside a coding or agent harness. In Hugging Face transformers, it's exposed as a reasoning_effort argument with named levels.

Why it matters: cost and latency are usually the binding constraints in production agents, not peak benchmark scores. Inkling's release sweeps its effort setting from 0.2 to 0.99 and shows the model reaching a given score at fewer tokens than rivals do at their default. The headline example: on Terminal Bench 2.1, Inkling matches NVIDIA's Nemotron 3 Ultra using roughly a third as many tokens.

Interestingly, this efficiency partly emerged during reinforcement learning. Over 30M+ RL rollouts, Inkling's chain of thought grew more concise on its own — dropping articles and connectives while staying comprehensible and reaching the same answer. Efficiency alone drove the compression; it wasn't a direct training target.

Inkling's benchmarks

Inkling is positioned as a competitive open-weights model rather than an outright leader. A few representative results from the release (run at effort 0.99):

• SWEBench Verified: 77.6% — ahead of Nemotron 3 Ultra (70.7%), roughly level with Kimi K2.6 (80.2%) and GLM 5.2 (80.0%), behind closed models like Claude Fable 5 (95.0%).
• Terminal Bench 2.1 (best harness): 63.8% — mid-pack among open weights; it trails GLM 5.2 (82.7%) here.
• GPQA Diamond: 87.2%; AIME 2026: 97.1% — strong reasoning.
• FORTRESS (Adversarial safety): 78.0% — the highest built-in safeguards of any open-weights model the lab compared, without over-refusing benign look-alike queries.
• Multimodal: 73.5% on MMMU Pro (vision), 91.4% on VoiceBench (audio) — among the stronger open-weights omni results.

The pattern is deliberate: broad and balanced across text, agentic, multimodal, and safety, rather than spiking on any single leaderboard. Independent verification will matter here — several scores rely on the lab's own or self-reported harnesses.

How it was built (and the distillation question)

A few architecture notes for the technically curious. The MoE design largely follows DeepSeek-V3 (256 routed experts, 2 shared, 6 active per token). Attention interleaves sliding-window and global layers at a 5:1 ratio, and — notably — Thinking Machines uses relative positional embeddings instead of RoPE, reporting better length extrapolation. Training used a hybrid Muon/Adam optimizer on NVIDIA GB300 NVL72 systems.

One honest caveat the company volunteers: to bootstrap post-training, it ran an initial supervised fine-tune on synthetic data generated by other open models, including Moonshot AI's Kimi K2.5 — a practice known as distillation. The lab says this was a small fraction of compute and that its next model will use fully self-contained post-training.

The speed claim is part of the pitch: OpenAI took about five years to bring its tech to market, and Anthropic roughly three; Thinking Machines says it did it in about nine months.

Why open-weights matters for AI agents

For teams building agents, the interesting shift isn't the parameter count — it's the deployment model. An open-weights base changes the economics: instead of paying per API call to a closed provider, you can run the model on your own infrastructure, fine-tune it on your domain, and keep the resulting model private. That's the same "own your workforce" logic driving interest in local, open agent stacks generally.

Inkling's controllable effort and agentic-harness training fit that world well. A model you can slow down for hard steps and speed up for cheap ones — inside your own agent loop — is exactly what long, multi-step workflows need. If you're weighing open options, our rundowns of DeepSeek V4 Pro, GLM-5.2, and Kimi K2.7 Code cover the other frontier-class open-weights models Inkling is measured against.

Do this with your own open-source AI workforce

Inkling's whole premise — that adaptable, self-hosted AI beats one-size-fits-all — is the same bet behind Eigent, the open-source "Cowork" desktop app that runs a multi-agent AI workforce locally. Where Inkling is a base model you fine-tune, Eigent is the workforce layer on top: a team of agents that automate real, multi-step workflows on your own machine, using the models you choose. If you want to put open-weights models to work instead of just reading benchmarks, download Eigent and start building your own agent workflows today.