The Hub had no blog post today. The day belongs to a new ternary Qwen3.8 family from Prism and two coding harness papers on Daily Papers.
On the Hub
prism-ml Ternary-Bonsai-2-27B
Prism-ml added Ternary-Bonsai-2-27B to the Hub. It is a 2-bit ternary of Qwen3.8-27B under Apache-2.0. The card went live on 2026-09-16 and was updated on 2026-09-17.1
The GGUF tree is 5.95 GB for PTQ1_0 and 7.21 GB for PQ2_0. Stock llama.cpp will not load these quantizations. The README points to the PrismML-Eng fork.
The card claims 84.78 average across 14 thinking-mode benches versus FP16. It claims 129.9 tok/s on RTX 5090 with PQ2_0 TG128 and 28.1 tok/s on M5 Pro with PQ2_0 TG128. The M5 Max 47.0 tok/s figure is pre-rotation and pending re-measure.
The family follows the Qwen3.6 Bonsai-27B line. Vision is optional via mmproj. Context is 262K.
NVIDIA SoL-Pi
NVIDIA's SoL-Pi paper 2609.20519 appeared on Daily Papers. It describes auto-research loops for coding harnesses. GitHub NVlabs/SoL-Pi is MIT with 2220 stars.23
The paper claims results on 51-task EdgeBench comparable to Pi on GPT-5.6 Sol and Opus 5. Token traffic is claimed 44.7 to 49.0 percent lower. API cost is claimed about one third. No Hub weights are linked.
The work focuses on action execution, context compaction, observation handling and delegated reading. It is a harness layer change, not a new model release.
Zoom harness design for coding agents
Zoom Communications paper 2609.20804 landed on Daily Papers. It is a 176-setting bake-off on SWE-Bench Verified and Terminal-Bench 2.1. No GitHub repo is linked.4
The paper varies planning, action space and context management across four models. Context management mostly stops overflow failures. Rule-based elision then summarization is claimed the efficiency winner.
Planning is claimed an accuracy scaffold for weaker models and a cost saver for stronger ones. Bash-capable models can skip predefined tools. The study isolates harness choices rather than shipping a new agent brand.
