Thinking Machines ships Inkling, a 975B open-weights MoE — what it means if you want a self-hostable frontier model
Mira Murati's Thinking Machines Lab released Inkling on July 15, 2026: a 975B-parameter, 41B-active multimodal Mixture-of-Experts with open weights, a 1M-token context, and a fine-tuning path on Tinker. Here's how it benchmarks against GLM 5.2, Kimi K2.6 and DeepSeek V4 Pro — and what a self-hostable open model changes for builders.
Thinking Machines Lab released Inkling, its first open-weights model, on July 15, 2026. It is the debut public model from the San Francisco lab founded in 2025 by former OpenAI CTO Mira Murati. (Source: Thinking Machines Lab)
Key facts:
- Inkling is a 975B-parameter Mixture-of-Experts model with 41B active parameters per token. (Source: Thinking Machines Lab)
- The weights are open, published on Hugging Face — including an NVFP4 checkpoint for NVIDIA Blackwell. (Source: Thinking Machines Lab)
- Context window is up to 1M tokens. (Source: Thinking Machines Lab)
- It is natively multimodal (text, images, audio, video), pretrained on 45 trillion tokens. (Source: Thinking Machines Lab)
- A smaller Inkling-Small (276B total / 12B active) shipped as a preview. (Source: Thinking Machines Lab)
The model uses 256 routed experts plus 2 shared experts, with 6 routed experts active per token, and relative position embeddings rather than RoPE. It exposes a controllable thinking effort dial from 0.2 to 0.99, letting you trade tokens for accuracy per request. (Source: Thinking Machines Lab)
What this means if you’re building with open models
Inkling is aimed squarely at builders who want a customizable model they can own, not just an API to rent. The lab pairs the open weights with fine-tuning on Tinker (its training API) and inference on partners including TogetherAI, Fireworks, Modal, Databricks and Baseten, plus self-host paths via SGLang, vLLM, llama.cpp and Hugging Face transformers. (Source: Thinking Machines Lab)
On coding, the numbers are competitive with the open-weights pack but short of the closed frontier:
- SWE-Bench Verified: 77.6% — near GLM 5.2 (80.0%) and DeepSeek V4 Pro (80.6%), below Claude Fable 5 (95.0%). (Source: Thinking Machines Lab)
- Terminal-Bench 2.1: 63.8% — behind GLM 5.2 (82.7%) and GPT-5.6 Sol (89.5%). (Source: Thinking Machines Lab)
- GPQA Diamond: 87.2%; AIME 2026: 97.1%. (Source: Thinking Machines Lab)
The pitch isn’t “beats the frontier” — it’s “a frontier-class MoE you can fine-tune and self-host.” The lab highlights effort scaling: Inkling reportedly matches Nemotron 3 Ultra on Terminal-Bench 2.1 at roughly one-third the tokens, which matters more for self-hosted inference cost than a leaderboard row. (Source: Thinking Machines Lab)
If you already run open coding models locally, Inkling slots in next to models like GLM 5.2 — see our guide on how to run GLM 5.2 locally for the vLLM/llama.cpp workflow, and our walkthrough for how to run NuExtract 3 locally if you’re building extraction pipelines on open weights.
The catch: at 975B total parameters, “self-hostable” means datacenter GPUs, not a laptop. For most teams the realistic entry point is Tinker fine-tuning or a partner endpoint, with local inference reserved for the NVFP4 Blackwell checkpoint or heavily quantized builds. Weights being open still matters — it removes vendor lock-in on the base model.
Sources
Source: Thinking Machines Lab