PrismML releases tiny reasoning LLM

PrismML released Bonsai 2 27B, compressing Alibaba’s Qwen3.8 27B to 5.9GB (≈9–10x memory reduction) while retaining about 98% of benchmark performance; the Caltech-founded startup has a $22.25 million seed round and investors including Khosla Ventures.

• Open-source engine Splash hits 144 tok/s on M5 Max2026-09-19
• Ternary Bonsai 2 (27B) released under 6GB2026-09-17
• Offload most of Qwen3.8 KV cache to RAM2026-09-16
• Swift‑Qwen3.8‑27B Cuts Overthinking Tokens2026-09-15
• Muse-glimmer-30b outperforms on creative benchmarks2026-09-11