Holo4-27B post-train cuts Qwen3.8-27B agent tokens by 79% on same task
solyarisoftware · x · 2026-09-29
H Company released Holo4-27B, a computer-use post-train of Qwen3.8-27B, and a benchmark using a self-playing Pac-Man game built in Godot shows a dramatic gap.
- Base Qwen3.8-27B: 197 agent calls, 11.4M tokens
- Holo4-27B: 68 calls, 2.4M tokens — 79% fewer tokens, 65% fewer calls on the same task
- The author notes accuracy remains to be verified, and an official 16.9GB Q4KM GGUF is available for local runs.
More from coding & agent
- Pydantic open-sources Monty: a Rust-based Python sandbox for running AI-generated code (8.4k stars) — samuelcolvin · 2026-09-29
- Training Agents IRL: GitHub and Hugging Face host Open Source AI Week event on multi-step agent RL — ben_burtenshaw · 2026-09-29
- eqsl-mcp: An MCP Server That Brings Natural Language to Amateur Radio QSL Logs — modelcontextprotocol · 2026-09-29
- After 75 interviews in 4 countries, one year proved 'AI can't code' wrong — realmeetjames · 2026-09-29
- 400 LLM Agents Living in an MMO Server: Lessons on Stale Perception, Fire-and-Forget Actions, and Load Shedding — kristiantalley679 · 2026-09-29
- Agent turns two NeurIPS papers into a 3-minute fully code-rendered launch film — liuziwei7 · 2026-09-29