2x RX 9060xt 16gb, is it worth it?
I'm planning to buy 2x RX 9060xt with 16gb each to run Qwen 3.6 27B and alike. Would it be a good investment? How much tk/s…
Latest
Tested which model can send best HTML email
Recently deployed https://github.com/Olib-AI/mailcue which comes with MCP server for managing emails. Wanted to see which model has better looking HTML email. The models I tested with…
I built a tool to turn your Claude Code sessions into fine-tuning data for local models
If you use Claude Code, every session is already sitting on disk as a .jsonl file under ~/.claude/projects/. It has real coding conversations: multi-turn edits, tool…
Full document redaction with Qwen 3.6 27B with a Pi agent harness
Link to full blog post with all method details, results, and links to all relevant code/skills/prompts at the bottom of this post. Apologies for not having…
Will Chinese Open Source Models be the only option soon?
US techbros do not just want to make money. They want total global control of everything. Releasing any more advanced AI interferes with that plan. submitted…
Model Registry: Torrents for open models using Hugging Face as a fallback web seed.
Hi, I created a repo/site for publishing/sharing .torrent files for popular open models, added web seed support and a few scripts to automate it. Repo: https://github.com/marella/modelregistry…
If it doesn’t make my PP better, I don’t want it
Highlights: 4 x 48GB modded 4090s 128GB DDR5 Pro WS WRX90E-SAGE SE 3000w PSU 240V/30A dryer line Q. Is putting a server on a dryer line…
Biggest model that is capable which can fit under 64 gb vram for the purpose of distillation
hi all, I have 64 gb VRAM, and I am looking for biggest model that I can use to distill prefer a reasoning model. even with…
Could AI game upscalers (like DLSS/FSR) benefit from lightweight game-specific adapters, especially for handheld gaming?
This is more of a thought experiment than a proposal. I couldn't find much discussion of this in publicly available papers, so I'm wondering whether this…
Does quantizing change the MTP draft rate?
Speculative decoding speeds up LLM generation by using a small "drafter" model to predict several tokens ahead of the main model. The main model then verifies…