How did we make DeepSeek outperform Opus

https://twitter.com/MrAhmadAwais/status/2050956678502420612

by ahmadawais • 15 days ago

34 7 15 days ago

ahmadawais4• 15 days ago

hey HN, sharing harness engineering deep dive on tool calling repairs for open models. i've been thinking about why "open model bad at tool calling" is almost always a harness problem, not a model problem. spent time looking at billions of tokens from DeepSeek (and other open models) in our coding agent. ended up building a tool-input repair layer on top of Zod. by the end, DeepSeek V4 Pro was beating Opus 4.7 in 6/10 of our internal evals. the main things that helped:

- most failures came from a small set of recurring schema mistakes - switched from preprocess-then-validate to validate-then-repair - handled some weird cases like markdown auto-links leaking into file paths

full writeup: https://x.com/MrAhmadAwais/status/2050956678502420612

video version (more detailed): https://www.youtube.com/watch?v=f61DCDwvFis

kadoban• 15 days ago

Thanks! Nice findings.

Is the output, that others can use, the research and findings, or is there a tool or something that came out of this that I can plug in to one of the common harnesses?

I guess that brief note in the twit is that it'll be part of that harness you're opening up in the future?

ahmadawais4• 15 days ago

Thanks! I've seen many developers use the entire tweet as a prompt to improve their harness. If you trace errors, you can find tool call failures per billion tokens. I've seen like many versions of this on pi, oc, or whatnot. This is already live in our hardness command code for almost two months now and we are now fixing one million tool repairs per one trillion tokens.

also decided to open source Command Code by v1 release, so hopefully end of the month. https://www.youtube.com/watch?v=laEzOCgtK6c (live stream on this)

maxjustus• 15 days ago

here's a lil pi extension I made yesterday based on it https://gist.github.com/maxjustus/3dd4c83865b1c439e26fdd925b...

bel8• 15 days ago

he trigger was watching deepseek-flash fail on the simplest /review run, every shellCommand and readFile call bouncing back with a raw zod issues blob, the model unable to recover because the error wasn't in a form it could read. by the end deepseek v4 pro was beating opus 4.7 6/10 times on our internal evals.

I think this is why Xiaomi forked OpenCode and created their own agent hardness, to reduce friction between model and harness:

https://github.com/XiaomiMiMo/MiMo-Code

ahmadawais4• 15 days ago

probably yep, i doubt they'll be able to maintain that well. we did a lot of work improving MiMo's cache thrashing, it was pretty bad early on. Now we're seeing 97% cache rate on their models after lots of prefix work.

How did we make DeepSeek outperform Opus

Related Stories