RynnUnify combines embodied understanding, visual planning and robot control in one 16B model

Embodied AI research leads the day, with RynnUnify for robot control and HiSTEVE 1.0 for Minecraft play. Nanbeige released a 33B model it says beats Qwen3.8-27B, and llama.cpp addressed an out-of-bounds write flaw.

1. RynnUnify, a 16B unified embodied model, combines embodied understanding with visual planning and robot control

   A single 16B model for understanding, planning and robot control reports 51.5% success on 20 composite RoboCasa365 tasks and a 46.7% average on three real-robot tasks.

2. HiSTEVE 1.0 hierarchical VLA agent plays Minecraft with keyboard and mouse

   A hierarchical vision-language-action agent handles long-horizon Minecraft tasks through keyboard and mouse, reporting 85.0% on MineStudio Bench.

3. Nanbeige releases Nanbeige4.3-33B on Hugging Face, claims edge over Qwen3.8-27B

   Nanbeige claims its 33B model beats Qwen3.8-27B on several benchmarks and plans to distill smaller versions, though only a community post links to the model page.

4. llama.cpp CVE-2026-116096 out-of-bounds write in Pocket-TTS component

   llama.cpp users should upgrade: a local out-of-bounds write via a malicious mmproj argument is addressed in build b11796, and b11871 aborts on that input.

5. TokenGate proposes token-gated attention for faster long-context prefill

   Researchers report TokenGate makes long-context prefill up to 11.3x faster.

6. IBM CEO says open weight models are now good and cheap enough for on-prem use

   IBM's chief executive argues open-weight models are now strong and cheap enough to run on company hardware, an opinion rather than a product launch.

7. RSIHub open-source framework standardizes comparison of agent-evolution methods

   Shared infrastructure for comparing agent-evolution methods such as GEPA finds pass@1 gains of up to 13.5 points that do not always transfer to held-out tasks.

8. UserArena benchmark tests shopping agents on real e-commerce user trajectories

   Built from 64,729 real sessions by 11,010 users, the benchmark grounds shopping agent tests in actual behavior and introduces OPTD to improve simulator fidelity.

Written by Claude (claude-opus-5-5) from the stories below.