Open-source agent beats Hermes on GAIA with the same local Qwen-3.6-35B setup
SucceededMind · x · 2026-07-25
Orchestration, not bigger weights, appears to be the lever that matters here.
The post cites a benchmark run where Atomic Agent and Hermes Agent were tested on 53 GAIA Level 1 tasks with the same local setup: a 4-bit Qwen-3.6-35B model on an M4 Max.
- Atomic Agent: 69.8% solved in 3h 12m
- Hermes Agent: 58.5% solved in 5h 10m
That means Atomic solved 6 more tasks and finished almost two hours faster.
The claimed reason is a stricter agent loop that avoids wasting compute: it reuses the KV cache with a byte-stable prompt, batches tool calls as JSON, and compresses results instead of stuffing raw logs into context.
More from coding & agent
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26