Thinking Machines ships Inkling 2 Mini: 26B looped MoE with NVFP4-native weights, ~20GB for laptops
testingcatalog · x · 2026-11-21
Thinking Machines released Inkling 2 Mini with open weights on Hugging Face and live on Tinker for inference and fine-tuning.
- Architecture: looped transformer, middle 12 of 24 layers run 3x per token; MoE with 26.4B total / 3.4B active params (5.4B activated per token counting loops); 20.1 AA Index, aimed at local agentic work.
- Precision: weights trained and deployed natively in NVFP4 (last four layers MXFP8), no quantization loss, 20GB total; the team ran it offline on a MacBook Pro mid-flight and built a playable flight simulator in minutes.
- Pruning and healing: started at sparsity 48 (8 of 384 experts), pruned two-thirds of experts midtraining to 128 per layer, then recovered performance with 1T tokens distilled from a larger internal Inkling model.
- Post-training: SFT followed by 2.2T-token RL over 2,000 steps with on-policy distillation.
- Ecosystem: day-one support from SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed; RL fine-tuning via SkyRL, Miles, NVIDIA AutoModel.
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