Tweet épinglé
Mateusz Mirkowski
3.8K posts
Mateusz Mirkowski
@llmdevguy
Autonomous agents, agentic engineering Building & testing agentic systems Exploring local LLMs
Remote work evangelist Inscrit le Mart 2013
149 Abonnements1.8K Abonnés
I would add just one thing.
Try to buy motherboard with 2 x8 pcie express ports. Cheap ones can give you one x16 and other x4 which will reduce speed greatly with two GPUs.
Lotto@LottoLabs
Keep it simple if you want to learn local ai 1) Build the cheapest rig that can house 1 used 3090 2) Who cares about ram, cpu, etc, buy ddr4 32gb, get a decent mid range ryzen/intel 3) download qwen 3.6 27b You’ll be into the rig like $1500 at most You won’t be able to upgrade much (maybe another 3090) but it’s a great starting point You don’t feel locked into learning about big rigs and upgrade paths If you like it then you can still keep and run the 3090 rig (trust me it’s useful) And you can build a new separate rig if you need more vram, on any platform after you have some experience and know what type of hardware you actually need, not what people recommend online!
English
🙏Please release a model that’s at least half as good as this meme.
Alexander Knigge@AlexanderKnigge
oh my god its happening @MistralAI has officially confirmed the upcoming release of Le Chaton Fat - 30T MoE with 256 experts - 1M context window - multimodal and multilingual - outperforms Fable 5 on every benchmark
English
Fine-tuned Super-Chaton Fat :
- 500T with 1B MoE
- 1B context
- 104.2% on SWE bench
- 98.2% on cute cat bench
Alexander Knigge@AlexanderKnigge
oh my god its happening @MistralAI has officially confirmed the upcoming release of Le Chaton Fat - 30T MoE with 256 experts - 1M context window - multimodal and multilingual - outperforms Fable 5 on every benchmark
English
@llmdevguy K2.7 seems to be around Sonnet 4.6 level.
GLM 5.2 (max) might be somewhere between Opus 4.6 and 4.7.
That said, GLM 5.2’s token consumption is seriously brutal.
English
🔥GLM 5.2 vs Kimi K2.7.
Which one is better?
Will test it soon.
What's your thoughts?
English
GLM-5.2 is now fully available for GLM Coding Plan users.
ZCode 3.0 is deeply optimized for GLM-5.2, bringing stronger Agent task execution, better long-context coding, and the new Goal feature for managing larger development objectives from planning to completion.
Coding Plan subscribers get 150% usage quota inside ZCode. New users get 5 days free with 5M tokens per day.
Download: zcode.z.ai
English
@maciekkukuczka Glm 5.2 tylko na z.ai na razie, ale pewnie na dniach będzie w open code go.
Polski
@llmdevguy No właśnie nie widzę jeszcze Glma 5.2 ani w openrouter, ani w zen ani w go. W każdym razie jak potestujesz to napisz opinię. Dzięki!
Polski
@maciekkukuczka Na kimi można bardziej polegać. Glm czasami bywa cholernie wolny. Jednak nowych wersji jeszcze nie testowałem. Polecam wziąć opencode go za 5 dolarów i się pobawić.
Polski
@llmdevguy Ok, a który z dwójki GLM/KIMI przy planie za okołostówkowym?
Polski
@maciekkukuczka Niestety m3 trochę nie dowozi. Poziom niżej od glm i kimi. Jednak dalej jest spoko do prostszych czy prywatnych zastosowań.
Polski
@llmdevguy To samo mi chodzi po głowie... No i jak się przy nich miewa MiniMax M3? Jakieś doświadczenia w .net (blazor)?
Polski
Tested DiffusionGemma 26B A4B vs Gemma4 26B A4B (BenchLocal, 4 bench packs):
Diffusion | Gemma4
> ToolCall-15: 83 | 97
> BugFind-15: 92 | 82
> DataExtract-15: 76 | 83
> ReasonMath-15: 50 | 89
Gemma4 leads overall, the ReasonMath gap is steep. But DiffusionGemma edged it out on BugFind-15, which surprised me.
Diffusion text quality looks rougher right now, but it's still an experiment. Curious to see where it lands long-term.
English
@redmonkeyAI I am testing it. It's tier below them unfortunately.
English
@arunoda True. I use 5.5 mostly. Kimi and glm are good enough in most cases but not in very complex tasks.
English
@llmdevguy I tried some really hard tasks yesterday with both. Still they are not Opus or GPT 5.5 level. Working with complex tasks involving backend, client & UI stuff they are not there yet. But for surgical tasks, it’s good. Both are better than their previous versions.
English
@llmdevguy I've only tried kimi so far for code review and it worked pretty well!
English
I hate that I am starting to like AI generated music better than other modern music.
There are some telltale signs music is AI generated, but it's subtle. My sister is big into making AI music also, so I've been forced to listen to quite a bunch this year.
Make it on a 3090.
English
K2.6 was the best Chinese model until today.
Now K2.7 is even better.
Amazing work.
Kimi.ai@Kimi_Moonshot
🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6. 🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates. ⚡️ 6x High-Speed Mode coming soon! 🔌 Available today via Kimi API and Kimi Code. 🔗 Kimi Code: kimi.com/code 🔗 API: platform.moonshot.ai
English
Google releases DiffusionGemma.✨
The new 26B-A4B diffusion text model runs locally on 18GB RAM.
It supports high-speed text generation, thinking, image, video and 256K context.
Run and train via Unsloth Studio.
GGUF: huggingface.co/unsloth/diffus…
Guide: unsloth.ai/docs/models/di…
Google Gemma@googlegemma
Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇
English