Give it a try!
🚨 JUST IN 🚨
Announcing the @NousResearch Hermes adapter for @Alibaba_Qwen Qwen3.8-27B!
Much more efficient and accurate agentic development with Qwen3.8-27B + Hermes!
- SWE-Bench Pro (hard 51) pass@1: 29.4% → 43.1% | with xhigh thinking: 45.1% → 51.0%
- Terminal-Bench 2.1 pass@1: 44.7% → 58.8% | with xhigh thinking: 58.8% → 57.6%
With thinking off, it matches the base model thinking at xhigh, while using 3.5-4.4x fewer tokens
huggingface.co/harness-hub/Q…
これめちゃくちゃ面白いな
Piハーネスに特化したQwen3.8 27BのためのLoraアダプタで、タスクあたりのツール呼び出し数が5〜6割くらい減るらしい...
主張するベンチだとthinking offでもめちゃくちゃ性能上がってるw
実現するなら個人でもできる範囲でGPU借りて好きなハーネス特化のLoraくらいなら作れそうだな!!!!
めちゃくちゃワクワクする!!!
↓モデルカードにあった学習深度によるモデルのツール呼び出し回数のグラフ
Announcing harness adapters for @Alibaba_Qwen Qwen3.8-27B on Pi harness!
We're releasing a LoRA adapter that makes Qwen3.8-27B a much better and efficient agent behind Pi harness.
SWE-Bench Pro (Hard 51 subset) pass@1: 29% -> 63%
Terminal-Bench 2.1 pass@1: 40% -> 56%
Much more efficient: 55-62% fewer tool calls per task 🤯
huggingface.co/harness-hub/Q…
The era of harness-specific adapters is now on! More harness adapters to come in the next days 🧵
Daniel retweeted
any ladies out there disgusted by the jensen huang cheating allegations?
if you own an RTX 5070 or higher, it’s time to take a stand
sell that disgusting piece of hardware to me for $50
don’t let that man’s GPU sit in your house. know your worth ladies
Daniel retweeted
Replying to @Saint_n0mad
And then you hit a weird bump and random lateral loads make the thing fucking explode into a million pieces.
Daniel retweeted
> need new machine quickly
> call the manufacturer, no answer
> email sales team
> next day they ask where we’re located
> Texas
> that’s X’s territory, here is his email
> email X
> auto responder, traveling and won’t be responding to emails
> call manufacturer, say I want to talk to different sales guy, they say I have to talk to X
> call X and he answers
> says he has machine in stock, great!
> says in stock means 4 week lead time
> how can in stock mean 4 week lead time I ask
> he says that’s the standard lead time
> I say any chance to make it faster?
> he says might be able to get it down to 2 weeks to ship out a machine that’s in stock
Daniel retweeted
I'm launching a site for free shader assets (CC0-licensed) for use in multiple engines: currently Unity, Godot, and Unreal Engine, with more effects and platforms planned in the future. cc0shaders.com/
Daniel retweeted
Is the Intel Arc Pro B70 slow for local LLMs? Plain decode, yeah. Turn speculative decoding on and it's a different card.
People keep asking me this since I ordered four of them for my own agents. Fair question, the raw numbers are rough.
One B70 running Qwen3.8 27B Q4_K_S in llama.cpp does 26.5 tok/s generating one token at a time. Two B70s on vLLM with no MTP did 26 tok/s on a 27B INT4. That's the number people screenshot and dunk on.
One token at a time is memory bandwidth bound, and the card's matrix engines mostly sit idle. Checking 4 to 8 drafted tokens in one step costs barely more than checking one, if the kernels are written for it.
Two receipts from this week.
budiony's llama.cpp PR moved the B70's multi-token path onto the XMX engines. Same card, same 27B, MTP drafting on:
Q4_K_S 38.7 to 51.0 tok/s
Q2_K 30.2 to 63.9 tok/s
A batch of 8 tokens on Q6_K went from 79 to 159.
That PR is closed without a merge right now, so don't look for it in a release yet.
@Daniel00Higgins shared a community run (credit JP-devv) of his Swift 27B on 2x B70, vLLM 0.30 XPU, TP2, MTP with 6 draft tokens:
120.5 tok/s single stream
328 tok/s across 4 streams
262K context, 1.28M token KV pool
about 500W for the pair
Same two cards, 26 tok/s without MTP and 120 with it.
Caveat, that run was a code prompt at temperature 0, which is the best case for drafting. Prose at a higher temp accepts fewer drafts and lands lower.
The B70 is a 32GB card with a lot of compute the software is only now learning to use. If you're judging it off a single-token llama-bench number you're judging the wrong thing.
Price is moving too. @zbian73 watched an ASRock B70 go from $1,400 to $1,945 on Newegg in a few days.
I haven't put numbers on my own four yet. When I do they go up here, MTP on and off.

ALT Table of Intel Arc Pro B70 tokens per second on Qwen3.8 27B. 1x B70 llama.cpp Q4_K_S: 26.5 no drafting, 38.7 MTP old kernels, 51.0 MTP best run. 1x B70 Q2_K: 37.6, 30.2, 63.9. 2x B70 vLLM 0.30 XPU TP2 INT4: 26 without MTP, 120.5 with MTP. Sources: llama.cpp PR 29864 by budiony; community run of @Daniel00Higgins' Swift 27B.
Daniel retweeted
Replying to @SawyerMerritt
Massive downgrade looks wise compared to the Roadster concept 😭
Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest 𝙻𝙵𝙼𝟸.𝟻 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses.
> 24B total parameters, 2B active per token
> Context expanded to 131K tokens
> A 128K-token vocabulary
> 33T pretraining tokens
> It is a reasoning model
> Very easy to fine-tune and customize
> Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.
> Open base and post-trained weights on @huggingface 🤗
1/
Daniel retweeted
Tibo promised 28 days of updates that would benefit “everyone.”
By Day 5, we get prompt prediction and Dots upgrades, both Pro-only.
DevDay was a slap in the face to Pro $200 users. Now it's Plus users' turn.
What a fucking joke. So much for “everyone.”
Daniel retweeted
🚨 PRICE ERROR
$2 Samsung Galaxy S25 on Amazon (Renewed)
amzlink.to/az0NxgUjhcwrJ
🤝 Paid partnership
Daniel retweeted
Denzel explica por qué usa la IA.
Un pequeña prueba explorando Minimax H3 en ComfyUI con mis nodos y métodos de inpainting.




























