Thinking Machines open-sources Inkling-2-Mini: looped MoE with 26B params, only ~5.4B effective active
Skyline34rGt · reddit · 2026-11-21
Mira Murati's Thinking Machines has released open weights for Inkling-2-Mini on Hugging Face, a general-purpose multimodal model (text + image in, text out) aimed at agentic/tool-use systems, coding assistants, chatbots, and RAG.
Key architecture details:
- 24-layer decoder-only transformer with sparse MoE feed-forward layers: 26.2B total params, 3.4B activated per token
- Core novelty is looped layers: the middle 12 layers run three times with shared weights (48 layer applications per token); each extra loop pass adds 1.0B effective active params, so 5.4B effective active per token
- Each token routes to 8 of 128 experts; attention mixes local and global layers; images are encoded via a ViT and processed jointly with text in a shared hidden space
- Per the team, "looping lets a model think more and write less," and the thought process remains monitorable at this scale; a variable-loop research model is also released
The model is designed to run on laptops, marking the first public artifact of the lab's looped-model research direction.
More from Models
- Tesla FSD 15's real-time view isn't 3D rendering — it's the AI model's own synthetic perception — mark_k · 2026-11-22
- Gemini Black Friday half-price: JP-region trick gets Pro annual plan for $91.54 — huangyun_122 · 2026-11-22
- Microsoft's USB4 RDMA lets mini PCs pool memory to run Qwen3-235B — JFPuget · 2026-11-22
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22