Unsloth ships local guide for Inkling 2 Mini: 26.4B MoE with 1M context, Apache 2.0
teortaxesTex · x · 2026-11-21
Thinking Machines Lab released Inkling 2 Mini, a 26.4B MoE (5.4B active) built on a Looped Transformers variant, Apache 2.0 licensed, with 1M context and vision support.
Key specs and scores:
- 48.2% on DeepSWE v1.1 and 77.9% on SWEBench Pro v2 (Hard) — strongest for its size, ahead of Qwen3.8-27B; ties Qwen3.8-27B on Terminal Bench 4.0 (5.6%)
- Unsloth Desktop updated accordingly: banded attention now priced correctly for longer context, plus Dspark detection for 2x faster decoding
- Recommended quant UD-Q4KXL at 17.6GB; 4-bit needs 17-18GB RAM/VRAM, lossless 8-bit needs 24GB; suggested settings: temperature 1.0, top-p 1.0, min-p 0
Also notable from teortaxesTex: looping models will likely integrate with speculative decoding — easy tokens are easy enough that loops aren't needed for them.
Related event: Thinking Machines Releases Inkling 2 Mini; Unsloth Ships GGUF for Local Use(4 posts)→
More from Infra
- SpaceX lenders push back on $40B of AI chips launched into space — flngr · 2026-11-22
- Homie: an offline Mac notch voice assistant running Qwen 3.5 4B locally — KungFuSaifooo · 2026-11-22
- AMD and PyTorch ship a playbook to run GPT-OSS-20B inference locally with ROCm — PyTorch · 2026-11-22
- Microsoft's full Outlook outage PIR confirms AI tooling wasn't to blame — aronchick · 2026-11-22
- User reports 400+ tok/s with Flare inference engine — capetorch · 2026-11-22
- Huawei Ascend-950DT quoted at 3x 950PR price; yield and SuperPod interconnect may explain the gap — teortaxesTex · 2026-11-22