Welcome Inkling by Thinking Machines

Source: HuggingFace Date: 2026-07-15 URL: https://huggingface.co/blog/thinkingmachines-inkling

Summary

Thinking Machines Lab shipped Inkling, its first model: a 975B-parameter (41B active) Mixture-of-Experts model natively handling text, image, and audio, with a 1M-token context window, released open-weight (Apache-2.0). Architecture is decoder-only with hybrid global/sliding-window attention, relative positional attention instead of RoPE, and discretized audio via mel spectrograms; day-zero support landed in transformers, SGLang, and llama.cpp, plus NVFP4/GGUF quantized variants. Self-reported benchmarks: 97.1% AIME 2026, 87.2% GPQA Diamond, 77.6% SWE-bench Verified, 73.3% MMMU Pro, 91.4% VoiceBench.

Implications

• Open-weight flood, marquee entry: the most-watched stealth lab of the cycle (Mira Murati’s Thinking Machines) chose to ship its debut model fully open rather than closed/API-gated — a data point against the “frontier labs stay closed” assumption.
• Independent verification pending: the SWE-bench Verified number is self-reported on Thinking Machines’ own harness; cross-lab coding leaderboards (GLM-5.2, DeepSeek V4 Pro) haven’t yet reproduced it independently — treat the coding-frontier claim as provisional until a common-harness comparison lands.
• Multimodal breadth as the real differentiator: native audio + vision + 1M context in one open checkpoint is a wider capability surface than most open coding-focused releases, positioning Inkling more as a general-purpose open frontier model than a coding specialist.
• Local-first inference: at 975B total params even NVFP4-quantized (~600GB) this sits well outside consumer/prosumer hardware — a cloud-tier open model, not a self-hostable one, unlike smaller efficiency-focused open releases.