← Back to feed

Thinking Machines Lab, the AI company founded by former OpenAI CTO Mira Murati, released Inkling on July 15 — a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, full weights under Apache 2.0, and a 1M-token context window.

The lab is explicit about the model’s market position from the first paragraph of its announcement: “Inkling is not the strongest overall model available today, open or closed.” The pitch is customization infrastructure, not benchmark supremacy.

Specs and Benchmarks

Inkling was pretrained on 45 trillion tokens spanning text, images, audio, and video — multimodal natively, not post-trained into it. Alongside the flagship, Thinking Machines is releasing Inkling-Small, a lighter variant with 12B active parameters trained on the same recipe, targeting inference-cost-sensitive deployments.

Benchmark results:

• SWE-bench Verified: 77.6% (bash-only harness)
• Arena Agentic Web Dev ELO: 1,258

The SWE-bench number sits at the mid-tier of the current open-weight field — below DeepSeek-V4-Pro (80.6%), Kimi K2.6 (80.2%), and Claude Sonnet 5 (85.2%), and above most 70B-class models. The Agentic Web Dev ELO of 1,258 is competitive but trails the frontier (Fable 5 holds the top of that leaderboard).

Inkling entered Arena’s Text Arena and Code Arena on July 16, and was added to the Agent Arena leaderboard on July 20.

Architecture Choices and Training Data

Inkling’s architecture mirrors DeepSeek-V3 — the Chinese open-weight model that redefined cost-efficiency when it launched in early 2026. The training process incorporated synthetic data from Chinese frontier models including Kimi K3. Both choices signal that the architecture and knowledge-transfer advantages that defined China’s open-weight push are now inputs to US-lab training pipelines, not barriers to them.

The irony is measurable in the benchmarks: Inkling trails the Chinese open-weight models it was trained on.

Fine-Tuning as the Core Product

The business model is built around the Tinker platform. Fine-tuning is available from day one, alongside an Inkling Playground for developer evaluation. Thinking Machines demonstrated the pitch in its launch materials: Inkling was shown fine-tuning itself — writing its own fine-tuning job, running it on Tinker, and evaluating the output. It used the OpenCode harness internally.

Inkling-Small at 12B active parameters is the production-deployment variant. At Apache 2.0, both models are commercially licensable without restriction.

Context

Thinking Machines Lab was founded in 2024. This is its first public model release. Murati departed OpenAI in late 2024 after nearly seven years including a tenure as CTO. The lab’s stated mission is AI that “extends human will and judgment” rather than replaces it — a framing that fits the emphasis on customization over raw capability.

For enterprises running open-weight fine-tuning pipelines, 975B at Apache 2.0 with managed fine-tuning on day one is a meaningful option. The Tinker integration narrows the gap between model acquisition and deployment that most open-weight labs leave to the user.