Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Kimi K2.7 Code is generally available in GitHub Copilot

github.blog

Post reply on HN
Print view View original

191–200 of 201 posts

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#191

Post by pbgcp2026 »

SwellJoe wrote: post #190
pbgcp2026 wrote: post #178

Earlier quoted context omitted.

A 4-bit quantization is useless in real projects.

The Gemma 4 QAT 4-bit models are basically indistinguishable from the 8-bit or 16-bit versions. I tested, and many benchmarks agree. For models where there isn't a QAT version, 6-bit is where it starts to degrade, but only a little if using dynamic quantizations like unsloth provides. When I use Qwen, I use the 8-bit quantization. But, when I use Gemma there is literally no reason to use anything other than the 4-bit…

I personally found in my work (long non-programming business-related context, 150k+) that only Gemma 4 31B full precision and full context works reasonably well. MoE Gemma 4, even non–quantised, makes clear logical mistakes or omissions. On real tasks. If I compare them on tiny tests – yes, sure, even i2 quants work great! :-)

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#192

Post by kamranjon »

MrDrMcCoy wrote: post #187
kamranjon wrote: post #144

Earlier quoted context omitted.

I'm really happy this is one of the top comments here, I am fully local as well. Just wanted to leave a note for folks who might not have the memory to run a big 32gb model - I just found out there are some pruned models that have really good performance and If I had a smaller machine I might try this pruned unsloth Q4 quant of GLM 4.7 flash that sits at 14gb: https://huggingface.co/unsloth/GLM-4.7-Flash-REAP-23B-A3B…

It's a good model for sure, but it's training set is based on some pretty old stuff. Would be good to post-train any languages / frameworks you care about being current into a LoRA for it. I even had to do that for the newer Qwen3.6, but it wasn't too hard thanks to Unsloth.

Curious what data sets you used?

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#193

Post by k__ »

russelg wrote: post #91
k__ wrote: post #51

Earlier quoted context omitted.

Yeah, v4 flash is dirt cheap, but it's running in circles quite often. Might very well be that a better model is cheaper if it gets things right the first try. Maybe I should route to a better model when v4flash hasn't solved after a specific number of tokens.

I'm having great success with DS4 Pro as my main model, while using DS4 Flash for subagents.

I tried DS4 Pro yesterday for one particular task that Flash struggled with.

That one Pro task cost $1.50 and the whole day of using Flash cost me $5.

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#194

Post by mellosouls »

theanonymousone wrote: post #138
mellosouls wrote: post #87

Earlier quoted context omitted.

where does it say that? its not available to me (also annual) at the moment via cloud but it said it is rolling out gradually, so I'm not too concerned. Tho I'm not overly excited either given Copilot pricing now; I reckon this should be at most 1x.

They say it here: https://docs.github.com/en/copilot/reference/copilot-billing... But then again they released MAI despite this, so I don't know.

Well I did check that first and didn't take the omission of Kimi to mean anything but the docs haven't been updated.

I guess it will be a wait and see thing.

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#195

Post by MrDrMcCoy »

kamranjon wrote: post #192
MrDrMcCoy wrote: post #187

Earlier quoted context omitted.

It's a good model for sure, but it's training set is based on some pretty old stuff. Would be good to post-train any languages / frameworks you care about being current into a LoRA for it. I even had to do that for the newer Qwen3.6, but it wasn't too hard thanks to Unsloth.

Curious what data sets you used?

Made my own from the latest docs of Ansible and Netbox for ops work. Even recent models seem to be trained on versions before some very breaking changes.

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#196

Post by selectodude »

MrDrMcCoy wrote: post #188
selectodude wrote: post #130

Earlier quoted context omitted.

Au contraire, I run models on WSL and my desktop reliably wakes up from sleep. Best of both worlds.

Genuine question: what are the good parts of Windows that keep you there? I've been such a happy gamer/coder/media editor on Linux for so long that I really don't understand how Microsoft isn't dead yet.

My desktop is a game console and the games I play work on windows. And my computer reliably wakes from sleep.

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#197

Post by MrDrMcCoy »

selectodude wrote: post #196
MrDrMcCoy wrote: post #188

Earlier quoted context omitted.

Genuine question: what are the good parts of Windows that keep you there? I've been such a happy gamer/coder/media editor on Linux for so long that I really don't understand how Microsoft isn't dead yet.

My desktop is a game console and the games I play work on windows. And my computer reliably wakes from sleep.

My Linux laptop and desktop both handle gaming and waking from sleep just fine. I'd be more surprised if they didn't. You might like CachyOS or Bazzite if you try it.

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#198

Post by Vinnl »

acpdev wrote: post #182
Vinnl wrote: post #101

Earlier quoted context omitted.

Ah OK, so the ACP connector ensures tool calls work with Zed, and communicates the available tools and their results to the harness, and then the harness mainly provides a system prompt and the API calls?

Harness does a lot of things potentially, but this depends on the harness. 1. Manages what is loaded in the context (I.e. sent to the model), this can be about compaction, summarization, removing things it judges no longer relevant, proving memory 2. Handling retries and how to resend data , lot of model specific behavior here 3. Wires up tools like web search or lsp servers 4. Skills and how to use them 5. Plugins/s…

Thanks! Although these, possibly with the exception of #2, sound like things that Zed handles, rather than the harness?

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#199

Post by acpdev »

Vinnl wrote: post #198
acpdev wrote: post #182

Earlier quoted context omitted.

Harness does a lot of things potentially, but this depends on the harness. 1. Manages what is loaded in the context (I.e. sent to the model), this can be about compaction, summarization, removing things it judges no longer relevant, proving memory 2. Handling retries and how to resend data , lot of model specific behavior here 3. Wires up tools like web search or lsp servers 4. Skills and how to use them 5. Plugins/s…

Thanks! Although these, possibly with the exception of #2, sound like things that Zed handles, rather than the harness?

No, not especially, the client (Zed in this case) can add some things (like @ing a file or class) and hint stuff but in all seriousness the harness is doing the things I’ve mentioned

You can see it from the different transcripts from the different tooling harnesses (´agents’ in agent client protocol speak) you can also just read the source for pi or zed or anything here don’t take my word for it.

You will see different web search from pi and codex, different tool calls, different context compaction strategies, etc.

The client has some control over this stuff but very little, like zed can pass in mcp servers to use or files to add to the context but the tooling harnesses can ignore those requests

Re: Kimi K2.7 Code is generally available in GitHub Copilot

#200

Post by k__ »

bel8 wrote: post #33
k__ wrote: post #27

Earlier quoted context omitted.

Nice. I paid $6 yesterday for DeepSeek V4 Flash on OpenRouter. That's like $120 dollar for a month, and it's not even a good model.

For DS4 it's much cheaper and reputable to use OpenCode Go $10/mo subscription, or directly with DeepSeek API.

[dead]
Post reply on HN