Vacuum 16T
https://huggingface.co/tsfrm/vacuum-16t A 16.5-trillion-parameter model that contains nothing. This model is just a ████ you to the labs and companies who say that "haha I have the…
Latest
How do you test your setup?
We all have been there, tinkering around with models is fun but we rarely do it with research precision and issues are often subtle and hard…
Single system with dual cards or two systems with single cards?
So I am in a conundrum and I'm thinking of asking for your opinion for the following: Currently, I have a 5800X3D gaming rig with a…
NEW Deepseek V4 Flash : MMLU-Pro , GPQA Diamond and truthfulQA ?
About the new deepseek v4 flash version / update, does anybody now about the new values about: MMLU-Pro GPQA Diamond TruthfulQA About the other values, its…
Xberg v1 is out
Hi all, I'm happy to announce that Xberg v1 is out. Xberg is the successor to Kreuzberg, equivalent to what would have been Kreuzberg v5. It's…
[Release] WinterMix — Qwen3.5-122B-A10B in native MLX: an 82 GiB build that beats 94–95 GiB quants, plus a 68 GiB build for agent swarms
TL;DR: I spent 9 days developing a new quantization method for MLX models and measured 18 variants against each other on a single M5 Max MacBook…
Setting up of a 16xGB10 (DGX Spark) cluster
Preparing this to be able to run locally frontier level open models. Deepseek v4 pro, Kimi K3, future ones like GLM 5.5 and Minimax M4. 16x…
Encrypted Clouds?
I love the progress happening on open models but I feel like it is kind of getting clear that hardware to run good sized models is…
DeepSeek-V4-Flash 284B on 5.3GB of memory
Following up on my Qwen 3.6 port, I wanted to keep adding models and ended up fixing a bunch of things along the way, so it's…
PSA for DeepSeek-V4-Flash-0731 users — don’t blow out your prompt cache with system role messages mid-conversation
DSv4F doesn't ship a jinja, but for distributions that do and faithfully reconstruct what DS releases in their chat template python, every system message is hoisted…