I post-trained Qwen3-Coder to fix bugs using an actual debugger. The result: Solve rate: 70% → 89% Median turns to fix: 46 → 19 (-59%) Instead of just reading code or print-debugging, it: - reasons from execution - inspects live variables and call stacks - sets breakpoints, steps, and... evaluates expressionsshow more

Across 27 heldout bug tasks, giving the base model the debugger harness alone didn’t help; without SFT + RL, solve rate was 67% (worse) and turns fell only 5%. The improvement came from training the model to use the debugger effectively.

Trained on @PrimeIntellect – more exciting results and in-depth blog post soon :)

Please bring to Qwen3.6-27B

the post-training recipe isn’t directly plug-and-play between these two: Qwen3.6-27B is a dense hybrid Gated DeltaNet/Gated Attention model, while Qwen3-Coder-30B-A3B is a sparse MoE.

@kylebrussell @grok explain this to me like im in highschool

Hey! Imagine an AI that writes code like a student staring at a textbook. This guy trained it to act like a real coder with a debugger—a tool that pauses the program, checks live variables (like peeking at what's in memory right now), sets breakpoints, and steps through line by line. Instead of guessing from reading code, it runs tests, investigates what's actually happening, finds the exact bug, fixes it, and re-tests. Result: Fixes bugs way faster and more accurately (70% → 89% success, half the steps). The video shows it debugging a tricky lock bug in Python. Super cool upgrade!

@ab198499 @moofeez @kylebrussell @grok FFS he meant the difference in training sparse MoE vs Gated DeltaNet/Attention model. But your answer was so ridiculous he just tuned out. And I don't care either. L

@ab198499 @moofeez @kylebrussell Ask Grok is currently available to Premium and Premium+ subscribers only. Subscribe to unlock this feature:

It’s interesting that learning to use the debugger is an underrated skill for most engineers, and it seems to be the same for ai too

+1, doesn't seem ideal that models currently spin on print-debugging loops

🔥

impressive work !

ty!! maybe worth adding to environments hub? 🤔

epic work!!

wouldn't have been possible without access to H100s on PI :)

@vincentweisser Love this work @moofeez! @vincentweisser what do we need to do to get Laguna on PI so we can see how it does here.

@moofeez would be epic - on it 🫡

Super cool! How much compute did this cost?

A few hundred dollars across synthetic data, evals, and training. imo worth it!

the comparison i'd want: same Qwen3-Coder with debugger access at inference, no post-training. splits the credit between 'has live state to reason from' and 'learned to use it well.' big chunk of the 19-pt jump might be the access alone.

@yourkaisensei this is exactly what I tested before going down the training rabbit hole giving the base model the debugger alone made the performance worse: I'll go into more detail here about the reasons behind this in my blog post.

this is cool and all but can it find nemo?

Excited to try this!

This is huge

would love the weights or training setup if you open source it

planning to open source everything 🫡

@willcb Did you give opus 4.6/ gpt5.5 a debugger? And see results there? Also did u try diff variations of giving the LLM a debugger?

great questions, I did run evals on Claude models towards the beginning of the project — the failure mode I observed was that the models would start a debug session but fail to use it effectively (shallow/incomplete debugger use), even on harder bugs not sure what you mean by “diff variations of giving the LLM a debugger”

@willcb Interesting. And by diff variations, im asking how you actually gave an LLM a debugger, like how did you explore it to the LM

@willcb I explored the standard harness + tool call approach, though there’s definitely room for experimentation here

@willcb Yeah exactly, probably more juice to squeeze by bitter lesson-ing the debugger in the harness. Looking forward to seeing your open source blog 👍

this is sick!

this is so sick 🔥

omg this is sick!!

let’s gooooooo 🥾🐛

good stuff!! did you ever have issues with the debugger going stale after a bunch of edits?

@botir33751732 the model learned to restart the debugger after making edits during RL :)

@botir33751732 though I also had some SFT examples with this behaviour

@trentster

you should do this for kernels and have it vibecode a parser for the nsight systems file format

jfc where does bro get the time

I put Claude on it

🔥🔥🔥

Did you create your own dataset for this? Have you considered porting to 3.6 now that it's killing the benchmarks?

This is interesting.

i don’t even think i can use a debugger fr lol

dw, my Qwen3-Coder finetune got you king

ok ! hugingface upload with the data you have pretrained ,to be transparent and see if the benchmark are real and not just pump

what confused me is that models don't even use the language servers, which is really strange?! Great work!

the absolute goat

wow man that’s actually pretty impressive nice work :)

🔥🔥🔥

How can I use this Mufeez?

model weights and harness will be available soon, you'll need to self-host 😃

@ClementDelangue this is sick very cool

Weights?

The 59% reduction in turns is the real metric here. Most debugging workflows assume infinite patience - but when you're context-switching between a sick kid and production fires, every extra iteration kills momentum. Teaching models to debug like experienced devs is huge.

Super cool idea.