Hacker News: pferdone

New comment by pferdone in "AfD Leak: What 421,179 Hacked Emails Show about the German Far Right"

pferdone — Tue, 20 Oct 2026 11:04:11 +0000

Different talking points. A party that wants to get rid of the system that enabled them to get voted for in the first place is anti-democratic. People need to be protected from that no matter who or how popular they are.

New comment by pferdone in "AfD Leak: What 421,179 Hacked Emails Show about the German Far Right"

pferdone — Tue, 20 Oct 2026 10:57:17 +0000

Which is also linked on the site

New comment by pferdone in "H3-metal – Native MiniMax-H3 inference for Apple Silicon"

pferdone — Tue, 11 Aug 2026 11:12:18 +0000

1) It would still run on the Mac's GPU.

2) Since it's unified memory, you won't have 96GB available.

3) I offered a solution that is usually recommended to the "gpu poor", if he's concerned with how much memory he would need.

4) I stated, that people already pointed out how he should be fine and that "gpu poor" doesn't apply to him.

5) "gpu poor" depends on what model you are trying to use. If you want to run Kimi or GLM you are still "gpu poor" even if you have an RTX Pro 6000 with 96GB of VRAM.

New comment by pferdone in "H3-metal – Native MiniMax-H3 inference for Apple Silicon"

pferdone — Tue, 11 Aug 2026 09:41:47 +0000

a friend told me there's a reddit called: unstable diffusion

New comment by pferdone in "H3-metal – Native MiniMax-H3 inference for Apple Silicon"

pferdone — Tue, 11 Aug 2026 09:39:50 +0000

you should have a look at https://github.com/deepbeepmeep/Wan2GP which is the goto tool for "gpu poor", although as people below already pointed out you should be fine with comfyui's standard setup aswell

New comment by pferdone in "OpenAI says its AI went rogue and launched 'unprecedented' cyber-attack"

pferdone — Wed, 22 Jul 2026 13:00:21 +0000

Public doesnt know about Huggingface. ChatGPT (OpenAI) says it‘s dangerous. They must know.

New comment by pferdone in "America pays workers just 27% of what its wealth allows – the worst in the OECD"

pferdone — Wed, 15 Jul 2026 10:45:57 +0000

You didn't even read your own article.

New comment by pferdone in "Steam Machine launches today"

pferdone — Tue, 23 Jun 2026 04:11:10 +0000

What does identity or sexuality have to do with it you fucking moron? These are people! People play games.

New comment by pferdone in "Gemma 4 12B: A unified, encoder-free multimodal model"

pferdone — Wed, 03 Jun 2026 17:29:48 +0000

But do I have the option to run it 'text only'?

New comment by pferdone in "Show HN: KVBoost – chunk-level KV cache reuse for HuggingFace, 5–48x faster TTFT"

pferdone — Fri, 22 May 2026 06:08:53 +0000

slop

New comment by pferdone in "The last six months in LLMs in five minutes"

pferdone — Tue, 19 May 2026 07:04:33 +0000

The consensus right now is that Qwen3.6 in its 27B and 35B-A3B versions is better for coding whereas Gemma4 is stronger when it comes to OCR, audio transcription and the likes. Margins are slim though and the harness at these model sizes is the most important factor.

New comment by pferdone in "ZAYA1-8B matches DeepSeek-R1 on math with less than 1B active parameters"

pferdone — Sat, 16 May 2026 18:40:55 +0000

That‘s so cool man! Congrats!

New comment by pferdone in "SANA-WM, a 2.6B open-source world model for 1-minute 720p video"

pferdone — Sat, 16 May 2026 13:14:29 +0000

First video with the guy walking the mountain in snow has consistency issues with the cave entrance. Which is "expected" at this model size?!

New comment by pferdone in "SANA-WM, a 2.6B open-source world model for 1-minute 720p video"

pferdone — Sat, 16 May 2026 13:12:27 +0000

Who wrote your comment?

New comment by pferdone in "ZAYA1-8B matches DeepSeek-R1 on math with less than 1B active parameters"

pferdone — Thu, 07 May 2026 11:31:24 +0000

I can see that and I don't know your setup, but there are people pushing >70t/s with MTP on a single 3090, with big contexts still >50t/s. 64k is not a lot for agentic coding, and IIRC 128k with turboquant and the likes should be possible for you. r/LocalLLM/ and r/LocalLLaMA/ are worth a visit IMO.

EDIT: just found this recipe repo, may wanna give it a go: https://github.com/noonghunna/club-3090

EDIT-2: this can also shave off a lot of context need for tool calling -> https://github.com/rtk-ai/rtk

New comment by pferdone in "I cancelled Claude: Token issues, declining quality, and poor support"

pferdone — Fri, 24 Apr 2026 16:44:57 +0000

pi.dev as well

New comment by pferdone in "Qwen3.6-Plus: Towards real world agents"

pferdone — Thu, 02 Apr 2026 15:04:27 +0000

> unlike almost all qwen models

Almost all means there have been ones before that were not open. So, no contradiction there.

New comment by pferdone in "Qwen3.6-Plus: Towards real world agents"

pferdone — Thu, 02 Apr 2026 14:53:34 +0000

They said in the last paragraph[0]:

"[...] In the coming days, we will also open-source smaller-scale variants, reaffirming our commitment to accessibility and community-driven innovation. [...]"

[0] https://qwen.ai/blog?id=qwen3.6#summary--future-work

What Did Ilya See?

pferdone — Fri, 06 Mar 2026 22:23:26 +0000

Article URL: https://www.youtube.com/watch?v=glWvwvhZkQ8

Comments URL: https://news.ycombinator.com/item?id=47281917

Points: 2

# Comments: 0

New comment by pferdone in "Z-Image: Powerful and highly efficient image generation model with 6B parameters"

pferdone — Sun, 07 Dec 2025 11:57:08 +0000

It‘s mainly due to system requirements that Flux.2-dev doesn’t get same usage as Z-Image. A 5090 needs about a minute to generate an image with a basic workflow with Flux.2-dev. But prompt adherence and scene/character consistency in edit mode is (way) ahead of Qwen-Edit-2509 if you ask me.