llama.cpp integrates Dots3 Note model, scoring 78.4 on SWE-bench Verified
victormustar · x · 2026-08-16
A PR submitted to llama.cpp adds support for the dots3-note-prev (280B/16B active) model. The key technical changes involve DSA (Dense-Sparse Attention) and SWA (Sliding Window Attention), requiring extensions to llama-kv-cache-dsa.
Benchmark results show strong performance across various tasks:
- SWE-bench Verified: 78.4 (vs Opus 4.8's 88.6)
- ARC-AGI-2: 81.4 (vs GPT-5.5's 85.0)
- LiveCodeBench v6: 91.5 (vs GPT-5.5's 96.2)
Logits validation has passed, with future plans to fix Metal matmul issues and add vision/audio support.
Related event: llama.cpp adds Dots3 Note model with SWE-bench 78.4 score(2 posts)→
More from Infra
- Running Qwen 2.5 on Mac Studio: Performance Tests and Agent Struggles — Over_Technology_1764 · 2026-08-16
- Nvidia reportedly in talks to invest up to $3B in SoftBank's SB Energy for OpenAI data center — Polymarket · 2026-08-16
- Exploring agent behavior when equipped with an owned knowledge library — dfinke · 2026-08-16
- Vietjet Invests Additional $9.5M in Starlink, Accelerating In-Flight Internet Deployment — XFreeze · 2026-08-16
- Weaviate adds test-time compute scaling to Search Mode, boosting retrieval performance significantly — dl_weekly · 2026-08-16
- New Book: Algorithms for Modern Hardware Open-Sourced on GitHub — thehiphopswami · 2026-08-16