PrismML released Bonsai 27B, a pair of low-bit quantized builds of Qwen3.6-27B aimed at on-device inference (MarkTechPost). The ternary variant stores weights as {-1, 0, +1} at 1.71 bits per weight in a 5.9GB file and retains 94.6 percent of the FP16 baseline across 15 benchmarks in thinking mode, while the 1-bit binary variant runs at 1.125 bits per weight in 3.9GB and holds 89.5 percent. Both ship under Apache 2.0 and are documented running on Apple Silicon, including M5 Pro and M5 Max machines and an iPhone 17 Pro Max, as well as on H100 GPUs (MarkTechPost).
Mistral AI released Robostral Navigate, an 8-billion-parameter model that takes RGB images and a plain-language instruction to move a robot through unfamiliar spaces without depth sensors or LiDAR (MarkTechPost). On the R2R-CE benchmark it reports a 76.6 percent success rate on the validation-unseen split and 79.4 percent on validation-seen, which the writeup says beats the best single-camera approach by 9.7 points and the best depth or multi-camera system by 4.5 points (MarkTechPost). Separately, a MarkTechPost analysis scored four coding agents, Mistral Vibe for Code, Claude Code, OpenAI Codex, and Cursor, on a single task of scaffolding a FastAPI endpoint through to a pull request, ranking Mistral Vibe first at 22 of 25, Claude Code and Codex tied at 21, and Cursor at 16; the comparison was based on documented features and vendor specifications rather than live runs (MarkTechPost).