JEVfire open-sourced: Qwen 0.8B clears Super Mario in-browser at 71ms per action
ricklamers · x · 2026-09-22
The open-source JEVfire project implements JEV/RLCD-inspired parallel decision-making for CUDA LLMs: typed variables are assigned from finite choices, independent fields are batched through vLLM for parallel execution while reusing shared context-prefix cache, and the model's existing LM head scores verified single-token labels mapped to allowed values, with JSON assembled in code.
- Headline demo: Super Mario recreated in-browser, played by a local Qwen 0.8B plus an explicit physics guard on an Apple M4 Max via WebLLM/WebGPU — 71.26ms mean worker inference, 511 accepted choices, level cleared in 40.12s (excluding model download and CPU forecasting)
- Ships with a vLLM API, game-agent examples, and reproducible benchmarks
More from coding & agent
- Agora paper: 13 LLM research agents self-organize via Git for 12 days — suchenzang · 2026-09-22
- Grok 4.7 drops, PSAISuite works unchanged — the case for stable LLM abstractions — dfinke · 2026-09-22
- Mistral: AI agents migrate 40,000 lines of Fortran 77 to C++ — dl_weekly · 2026-09-22
- One-line input to finished episode: an agent pipeline built on CREAO — socialwithaayan · 2026-09-22
- UiPath: Haiku's agent success rate halves with every doubling of task length vs Sonnet — EvalRaccoonDev · 2026-09-22
- Storewake MCP server puts App Store rankings, RevenueCat revenue and Apple Ads into your AI assistant for $19/mo — Pfernan95 · 2026-09-22