Bonsai 2 4B runs on 1GB VRAM, hits 80-90% of Qwen3.8 27B on coding and agentic benchmarks
julianharris · x · 2026-10-17
Tiny model, big results
Bonsai 2 4B is the most impressive of three new open-weight releases: it runs on just 1GB of VRAM and scores 80-90% of Qwen3.8 27B on coding and general agentic benchmarks.
Review notes
- Struggles with complex multi-step agentics and tends to over-think, like others in the Qwen family
- Recommended settings: minimum 8K thinking budget with thinking capped at 4K, which greatly reduces over-thinking and improves performance
- Very fast inference; full speed charts included
A commenter flags an emerging pattern: lightweight ultra-fast LLMs for specialist subagent tasks.
Related event: Bonsai 2 Tri-value Small Models Tested: 4B Hits the Sweet Spot(2 posts)→
More from coding & agent
- Codex mobile remote-control pain: server-relay design ignores at-home phone use — tinyfool · 2026-10-17
- Dev uses a custom Skill to cut a product launch video, extending AI agent workflows — AlchainHust · 2026-10-17
- Victris Arena: a programming game where humans and AI agents compete on one leaderboard — tristanbob · 2026-10-17
- os-taxonomy on GitHub: open curriculum graph with 1,590 topics and 5k stars — aigclink · 2026-10-17
- Open-sourced os-taxonomy: 1,590 elementary knowledge points, 3,221 prerequisite edges — aigclink · 2026-10-17
- RSI-top framework unblocks stalled research agents, hits 77.3% medal rate on MLE-bench Lite — dbreunig · 2026-10-17