Qwen3.8-27B at 1.8bpw collapses to 13% on MiMo RL bench, EXL3 tests show
casper_hansen_ · x · 2026-10-24
casperhansen tested EXL3-quantized Qwen3.8-27B on MiMo RL environments:
- BF16: 63.2%
- 2 bpw (v1.4.2): 55.5%
- 1.8 bpw (v1.4.2): 2% (near-total collapse)
- 1.8 bpw (v1.6.4): back up to 13.2%
He notes that at 1.8 bpw roughly 27% of weights get pushed to 1-bit, which is what causes the collapse. A useful data point on the practical floor of extreme quantization for large models on low-VRAM hardware.
More from Models
- 1,626-question test: free e5-large embeddings beat OpenAI's paid tier in German and Spanish — julsimon · 2026-10-24
- Muse voice assistant insists the fake "iPhone Duo" doesn't exist — jesselyu · 2026-10-24
- SWE-sweep Leaderboard: Top Models Double Bug Finds at Half the Cost, Still Under 10% — stanfordnlp · 2026-10-24
- AI Finds Software Flaws Faster Than Fix: 45,207 CVEs Logged by Late July — rohanpaul_ai · 2026-10-24
- Kimi K3.1 Pro launches on mainland China site while overseas users still capped at K3 — teortaxesTex · 2026-10-24
- 156MB Phonon-2.5 Beats Whisper large-v3 on Multilingual ASR at 1/20th the Size — julianweisser · 2026-10-24