benchmaxxed
the too good to be true alarms are going off. has to be another benchmaxxed or misleading model. I doubt it has the specs claimed.
it's an extremely impressive agentic model for its size, I have been using it
this model is impressive. I use Qwen 3.6 35b-a3b and Qwen 3.6 27b every day, and I can tell you this model is here to stay. It beats the 35b handsdown, but is now also progressing through 14-stage workflows that have been challenging for the 27b. I'm now putting it up to long-term challenges.
this model is impressive. I use Qwen 3.6 35b-a3b and Qwen 3.6 27b every day, and I can tell you this model is here to stay. It beats the 35b handsdown, but is now also progressing through 14-stage workflows that have been challenging for the 27b. I'm now putting it up to long-term challenges.
I hope you update us about the results of your tests. I hope this model really is something worth using.
What about comparison with Qwen's own AgentWorld, which was released like two days prior? It is identically insanely good at toolcalls, and has not had a miss on anything thrown at it on my end. Is this a secondary party's implementation of what Qwen (apparently) set out to build with AgentWorld? Both look insanely good, but I have only tried AW and not AA1... apparently models trained on simulating a terminal-output aware environment is such a game changer. As even AW under a harness of any sort is like 10 lightyears ahead of what we just had a few months ago with normal (non-agent) models. Actually incredibly impressive for Hermes/Openclaw/Pis/etc
What about comparison with Qwen's own AgentWorld, which was released like two days prior? It is identically insanely good at toolcalls, and has not had a miss on anything thrown at it on my end. Is this a secondary party's implementation of what Qwen (apparently) set out to build with AgentWorld? Both look insanely good, but I have only tried AW and not AA1... apparently models trained on simulating a terminal-output aware environment is such a game changer. As even AW under a harness of any sort is like 10 lightyears ahead of what we just had a few months ago with normal (non-agent) models. Actually incredibly impressive for Hermes/Openclaw/Pis/etc
Thanks for the attention! I also think AgentWorld is very impressive. My understanding is that AgentWorld starts from the idea of training models with simulated real-world feedback, and surprisingly obtains a base model with quite strong agentic capabilities.
Agents-A1 is a bit different: as we mentioned in our technical report, it is post-trained based on Qwen3.5, and is mainly optimized for agentic search and scientific reasoning tasks.
Both directions are very exciting, and I’d love to hear your real hands-on feedback if you get a chance to try Agents-A1.
Thanks for the attention! I also think AgentWorld is very impressive. My understanding is that AgentWorld starts from the idea of training models with simulated real-world feedback, and surprisingly obtains a base model with quite strong agentic capabilities.
Agents-A1 is a bit different: as we mentioned in our technical report, it is post-trained based on Qwen3.5, and is mainly optimized for agentic search and scientific reasoning tasks.
Both directions are very exciting, and I’d love to hear your real hands-on feedback if you get a chance to try Agents-A1.
A few things actually. I know the realm of local agents is rather niche for consumers due to constraints of hardware... but seriously. If one has a large consumer card like a 4090/5080/5090... this is the model for you.
I'm currently running the int4 on consumer Blackwell with compute1/Agents-A1-GPTQ-INT4-Sym — this addresses the large problem consumer cards on vllm/sglang have (Marlin MoE vs Triton) — and I know it doesn't sound like of much importance, but here's the kicker:
With this integer quant (I'd prefer floating-point or even nvfp - but I am too lazy to get it working as of now), this model hits an average of 120-160 tok/s at times, with hardly any falloff at long context on vllm. (none that I could notice, anyway). It excels at toolcalls. In fact, Agents A1 is significantly handling browser-use and tool use in all of the tests I throw at it. Plugging directly into Hermes, which is wired with loads of basic MCPs any generic Agent might need - this model fails at none of it. It handles modern web MCP tooling perfectly, as well as writing extremely grounded code.. it also fixes itself.
The comparison to AgentWorld is where I really have curiosity. Everyone insists upon (also) Qwen 27B for code, and that is where they may have their only benefit. 35B MoE is not far behind. I'd like to see these critics get their 27B dense models to handle failed toolcalls or even basics of browser control. Agents A1 does all of this out of the box, including going above and beyond for the scientific reasoning tasks with websearch, and natively handles the linux ecosystem like it's a direct appendage of itself.
The weights make it possible for consumers. The MoE architecture makes it usable for speed. And the training here is significantly more grounded than AgentWorld in curated dataset adherence. I really appreciate what your org does, and I'm excited to see where you guys move next.