Model is decent, its faster than Qwen and pretty close on real world tool calls
It's pretty good. It will overthink sometimes, but I didn't necessarily get anything ever stuck in a thinking loop. Overall 2-3x faster than Qwopus. It will talk more between tool calls and will sometimes pause to verify intent more than Qwopus.
Okay it can use tools... can it actually complete an end to end task?
I asked it to create a client and an onsite for today at 10am with 5 different line items between parts and labor as well as a description/notes section.
It created the Client, Created the onsite properly, filled in the onsite data, then asked if I wanted to create a stripe invoice from it and send the email. So yes, I would say it completed the task.
Okay it can use tools... can it actually complete an end to end task?
Nope, not on anything challenging, it'll collapse into reasoning loops.
It doesn't seem to work well at all with Github copilot extension in vscode. Also not able to get it to think on complex problems.
I have been using it as a daily driver in place of 27B (locally) and glm 5.2 (from an endpoint) since it dropped. Running on a dgx spark, the nvfp4 version, with repetition penalty 1.15. Hasn't looped since upping the penalty (i did see it loop once on even 1.1 ). In terms of 'artificial' benches, it outdid both 27B and ds v4 flash on ML engineering tasks, it was horrid on web design (which doesn't matter to me personally, but it was an interesting failure mode in my testing). In godot it outdid 27B but ds v4 flash had a lot less issues than the other 2. It seems very good with complex algorithms and architectures. In terms of 'real' work, I've used it a bit for assistance for work (ASR Model Engineering and training framework work) and it performs very well. It can absolutely handle non-design work, if it doesn't enter an actual loop (on occasion it looks like it's looping, but it leaves the loop on its own). I hope the looping thing gets resolved with an extra checkpoint, as this model feels special, and I'd rather use it without a repetition penalty if possible without worrying of a potential loop after 200k ctx.