Entropix creator _xjdr says Moonshot's Kimi K2.7-Code outperforms K2.6 while cutting daily token usage to 592 million

here is a snapshot of a single repo, but k2.7-Code has been much more terse (much much more) without a perceived drop in quality (in fact id say mostly an overall increase)

same repo, same workloads, some work done in parallel for evals

@keennay you'd be surprised how much intelligence is in the harness and i've spent a lot of time making my claude code fork with well with k2.* family.

@xlr8harder @keennay what you call out is one of the reasons i said github was the second largest risk behind anthropic / claude code

@teortaxesTex this is hooked into my inference system so it would be hard, but if i have time i'd love to figure out how. so far, i've been really impressed by k2.7. im probably going to abandon my k2.6 FT and focus on optimizing for 2.7

@_xjdr have you ever used OpenCode → your local Kimi inference for your workload?

@_xjdr @keennay @secemp9 Oh I misunderstood at first. You're thinking takedown resistance?

@_xjdr @keennay Not sure if it's the right basis for you but @secemp9 has a python port of pi that i've been working with

@_xjdr Can you check Composer? I don't get why it's so much cheaper

@_xjdr @keennay any thoughts on open sourcing this? If it's a fork I guess you can't, but perhaps it could be ported to another framework?

@joseluissaorin its hard to say because i use these models almost entirely for code and engineering so i would be unable to properly gauge how robust the pretraining is. for my purposes, it is excellent

@_xjdr how custom is "custom" here? like swapping model calls or is there actual orchestration around the code generation step

@_xjdr Do you feel it is far off in terms of pretraining? How do you feel about its "creativity" or deep knowledge and memorization?

@_xjdr k2.7 being its own thing is the correct take. every model that tries to be opus clones itself into mediocrity. moonshot actually shipped something with a personality for once.

@_xjdr not sure beating a harness-specific k2.6 means 2.7-Code is genuinely stronger could just mean the fine-tune drifted from your current config. saw that pattern with a tool setup i optimized then updated. did your harness change much since you trained k2.6?

@_xjdr This is a pretty insane ratio

@bygregorr i didn't say it was stronger, simply that it worked surprisingly well out of the box and in my initial testing compared favorably to what i have been using . these models are good enough now that 'stronger' is typically reductive. most have strengths and weaknesses

@_xjdr @keennay Sounds pretty cool and interesting: Are you dogfooding your training stack for harness post-training or have you written or talked about what you’ve done anywhere?

Is it just SFT to get it to know tool formats or full RL?

@ApplyWiseAi i have both, im not sure what would be practical to OSS . i.e, specific k8s tooling and flows for managing gb300nvl72s is probably not the most practical to support.

@_xjdr thank you mr oracle, this will heal the scars fable has left us

@_xjdr cool teardown. for anyone in this thread still on fable watch, leaving this here