Seti @ Home

[Seti @ Home @seti_home]

@aijoey @Alibaba_Qwen @QwenDevs May you share the prompt as well.

[Seti @ Home @seti_home]

@seanvishwakarma I'm using Halogen Flash Server. https://t.co/3HVG3LVw4Y

[Seti @ Home @seti_home]

5 coding agents. Same Halloween pixel-art prompt. Same local Qwen3.8 Flash Next (Q4XL→HGN) on Halogen/Ubuntu.

Results:
Claude Code: 14m 42s
OpenCode: 36m 30s
Copilot: 58m 30s
Pi: 1h 20m
Codex: 4h 4m

16.6× difference in completion time. Which looks best?

[Samuel Zeng @SamuelZengML]

Introducing Edge0-145B and Edge0-120B, @Edge0_AI’s first device-native 100B-class foundation models with open weights.

Cloud-level intelligence on your laptop. Free, offline, and fast enough for real vibe coding.

- Smart: beats Claude Opus 4.6 on Terminal-Bench 2.1 (86.5% vs. 78.2%)
- Light: needs just 29G memory, 1/3 of a standard setup (100G+)
- Fast: decodes 75+ toks/s, prefills 1050+ toks/s on RTX Spark and Mac.
- Steady: handles 128K–256K tokens of context without slowing down, even on long tasks

Frontier-class agentic coding. No cloud API. No token bill. Your code and data never leave your machine.

Open weights are live on @huggingface. Run it locally or plug it into your favorite coding agent.

[Seti @ Home @seti_home]

@LatentOperatorX DeepSeek is suppppeeerrrr slooooww

[Dennis Rye @Dennis_Rye]

This is what TensorFold 1.2.7 looks like on AMD Strix Halo, and this is what my local dev branch looks like.

There's still more performance available on AMD.

[Seti @ Home @seti_home]

@Rainmaker1973 Lesson from this: Always get something for your woman and you won't regret it.

[Can Bölük @can1357]

"oh no what did we regress from pi"
> 8 tasks failed because reward hacking counts as failure (we add web search, duh)
> 3 tasks failed because we keep bg tasks scoped to the session so that people don't have runaway scripts
jebaited once again 💀

[Seti @ Home @seti_home]

@joelniklaus @deepseek_ai @huggingface Could you try copilot desktop harness?

[Seti @ Home @seti_home]

@coldniko Soon soon final.doc

[Seti @ Home @seti_home]

@joelniklaus @huggingface That's a quite jump

[Seti @ Home @seti_home]

@halogenengine Is it possible for Windows support for Halogen?

[Seti @ Home @seti_home]

@academind_real But then, how we will learn new skills? Are we going to do tldr every time when we need to understand something new.?

[[archer] @archer_uwu]

this is whats going to happen to epic games in about a year

[Seti @ Home @seti_home]

@halogenengine @AIatAMD Is it possible for Windows version of halogen?

[Seti @ Home @seti_home]

@loktar00 May you please share the prompts as well. I want to compare your results with Qwen 3.8 flash next for example

[Seti @ Home @seti_home]

@joelniklaus @Thom_Wolf Try with copilot as well. It looks good and fast

[Jules Horn @juleshorn01]

Do it for 21 days daily, come back and tell me what you feel. Thank yourself later 🙌

[Casper Hansen @casper_hansen_]

DeepSeek v4.1 Flash 1017 at home at 15 tokens/s on my RTX 3090 build is super viable now.

My roadmap:
1. Build my own inference engine, then
2. Undetected browsing plugin, then
3. New quantization framework
4. RSI over step 1-3 at home

You can do the same for $2.5k..

[Loktar 🇺🇸 @loktar00]

Some people were asking for the prompt for this, it was pretty straight forward, throw it into a harness and have it continue as much as you want.

Prompt below.