Image: uni-robot-agent
Qwen3.8-Flash-Next can operate a robot arm without any robot-specific training, using an open-source harness built by Qwen developers at Tsinghua and UC Berkeley. On 19 real tabletop tasks, the setup scored 49.50, slightly above the 47.75 of a fine-tuned π0.5 model, the team says.
The harness, uni-robot-agent, connects a general model to robot tools through ordinary tool calls. The model gets perception, motion, skills and termination tools, and it does not need new weights. The repository is MIT licensed and ships a CLI, Python SDK and local dashboard for Codex and Claude Code runtimes.
Well, turns out Qwen3.8-Flash-Next can do more than work with software tools, it can operate real robots too. A group of Qwen developers from Tsinghua and UC Berkeley recently built uni-robot-agent, an open-source harness that connects general-purpose LLMs to robotic systems. It gives Qwen a way to interact with the physical world through tool calls without changing the model weights. With this harness, Qwen3.8-Flash-Next reached the performance level of some previously reported closed-model robotic agent systems. The average API cost was $1.09 per task, compared with $21.29 for GPT-6 Astra. On 19 real-world robot manipulation tasks. Without any robot-specific fine-tuning, Qwen achieved a slightly higher score than the fine-tuned π0.5 The harness is open-source if anyone wants to tinker with it.
— Qwen Developers (@QwenDevs) November 3, 2026
In simulation, Qwen3.8-Flash-Next scored 31.39 on RoboDojo, above the published 28.97 for GPT-6 Astra with RoboProbe. The paper says this is the first system to beat closed-weight foundation models on that benchmark using open weights, without privileged inputs or distillation from those models.
The real-world result needs more context. On the same 19-task suite, GPT-6 Astra scored 75.50, far above Qwen's 49.50. So the open model matches a fine-tuned robot policy and beats Astra in simulation, but it does not match Astra on the physical tasks the paper tested.
Image: uni-robot-agent
Cost is the stronger claim. The team estimates $1.09 per task for Qwen3.8-Flash-Next against $21.29 for GPT-6 Astra, roughly 19.5 times less. Astra still scores higher. The paper frames the result as a cheaper trade-off, not a straight win.
The public release has a gap. Version 0.1.0 includes simulator integrations for RoboDojo, RoboUse and RoboLab, but the README says real-robot connections for ARX One and Franka are "coming soon." The 19-task evaluation ran on in-house hardware, so outside teams cannot yet reproduce it with the published package.
Qwen3.8-Flash-Next is a 125B mixture-of-experts model with 6B active parameters and open weights on ModelScope and Hugging Face. The harness is the interesting part: it turns robot control into a tool-use problem that any strong model can attempt. That lowers the barrier to testing robot agents, but the physical integration and independent benchmarks are still missing.
Comments