Community releases Groot MoE family: 4 models up to 280B, claiming 63.47 on SWE-bench Pro
AdinaYakup · x · 2026-10-27
Community team GrootTeam published the System-Two collection on Hugging Face: four MoE models (19B/A3B, 65B/A6B, 238B/A18B, 280B/A16B) with FP8/NVFP4/GGUF variants.
- 19B/A3B: claims 128K context on a single 16GB GPU, outperforming larger models on reported benchmarks
- 65B/A6B: claims 28% fewer output tokens vs Qwen3.8 Flash Next
- 238B/A18B: claims wins over larger models on 6 agentic benchmarks, 1M context
- 280B/A16B: claims 63.47 on SWE-bench Pro and $0.13/task on DeepSWE 1.1
All numbers are self-reported and unverified; the models are still niche (hundreds of downloads).
Related event: GrootTeam Releases Open MoE Models from 19B to 280B(2 posts)→
More from Models
- Blogger: Apple's AI will still be terrible even on iOS 270 — lipeng0820 · 2026-10-28
- SlopCodeBench October report: Astra beats Opus 5.5, xhigh effort deemed harmful — teropa · 2026-10-28
- OpenBMB ships 1.3B OCR model scoring 93.68 plus a 2B AI judge for OCR errors — Equivalent-Grass-527 · 2026-10-28
- Fangcun AI publishes third-party safety evaluation of DeepSeek V4.1 Flash-1017 for manipulation risks — pstAsiatech · 2026-10-28
- NVIDIA open-sources a LoRA post-training workflow to localize Nemotron with synthetic personas — NVIDIAAI · 2026-10-28
- Carbon-X hits frontier-level DNA understanding: Qwen3-14B finetune scores 66.7%, near GPT-6.1 — lvwerra · 2026-10-28