FULL STORY
PrismML's Ternary 27B Claims Meet Community Benchmarks
PrismML claimed its 5.9GB ternary-quantized 27B model retained 98.2% of full-precision scores, but community benchmarks on 16GB GPUs found Qwen's IQ3_XXS quantization roughly 3x faster.
2026-09-19 ~ 2026-09-20 · 2 episodes · 4 posts
Episode 1 · PrismML's 98.2% Claim for Ternary 27B Model Challenged in Real-World Tests (2026-09-19, 2 posts)
PrismML claimed its 5.9GB ternary-quantized Qwen3.8 27B retains 98.2% of full-precision performance, but community testing showed the score came from cherry-picked static benchmarks and the model failed badly on real agent tasks.
- Qwen3.8 27B's 98.2% Benchmark Slammed: Real Agentic Tasks Fall Apart — teortaxesTex · 2026-09-19
- Real agent tests expose gap between 98.2% benchmark scores and actual quantized model performance — TheMoonMidas · 2026-09-19
Episode 2 · Benchmarks: Qwen 27B IQ3 quantization runs 3x faster than Bonsai ternary on 16GB VRAM (2026-09-19, 2 posts)
Community benchmarks on 16GB VRAM compared Qwen3 27B IQ3XXS quantization against the ternary Ternary-Bonsai-2-27B-PQ20, finding the Qwen variant roughly 3x faster while both heavily-quantized 27B models remain functional for local use.
- Qwen 27B IQ3_XXS beats Bonsai ternary PQ2 by 3x on 16GB VRAM local test — Danmoreng · 2026-09-19
- Community Benchmarks: Ternary-Bonsai-2-27B Extreme Quantization Is Not Lobotomized — Fancy-Snow7 · 2026-09-20