TrendPulse

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 calls, reasoning traces. That's training data you already generated for free. The problem is the format is not what any fine-tuning framework expects. So I built claude_converter to bridge that gap. What it does: Converts Claude Code .jsonl sessions into the messages format that apply_chat_template() consumes d...

Tried creating a quick 3d game with it, after 3 prompts, it got me this(checkvideo). If I compare this with qwen3.5-35b-a3b, it was not able to successfully generate this and was failing even after multiple prompts. Harness: Claude Code How is your experience so far ? https://reddit.com/link/1uh8von/video/csrnhpwy2v9h1/player

Highlights: 4 x 48GB modded 4090s - 192GB VRAM 128GB DDR5 Pro WS WRX90E-SAGE SE 3000w PSU 240V/30A dryer line Q. Is putting a server on a dryer line a good idea? A. No, or emphatically yes. Splitters on this line are not code compliant, so I have to turn off the server to use the dryer, OR buy a smaller dryer that can go on the 20A. Also I've had two nuisance trips while idle in the past month due to laundry GFCI. A dual conversion pure sine wave UPS is on the way. This room is...

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 Site: https://modelregistry.io Web seed will be used as a fallback to download files from Hugging Face if no peers are available. To make this work ( https://www.bittorrent.org/beps/bep_0019.html ), I had to build a small backend service that redirects the requests from BitTorrent clients to the...

Too many times I hear people whine about not being ble to run SOTA models or claim it would require $50k, or $100k. https://www.ebay.com/itm/398079051468 Epcy Motherboard & CPU - $460 https://www.ebay.com/itm/206374955959 P40 24gb - $230 get 2 - $460 https://www.ebay.com/itm/318489798853 512gb dd4 $1000 Total = $1920. You need PSU, Storage, Fan for P40. You can source those for $350 easily. But let's go ahead and budget $580 to put the total for everything at $2500. You ca...

Sunday experiment. Same prompt to both. Build a voxel world in plain C. No engine, no game library, no framework, just the compiler. The model does its own chunk meshing, render loop and memory management by hand. Left is Claude Code on Opus 4.8. Right is Qwen3.6 27B local on vLLM, the new NVFP4 quant, 256k context. Runs around 130 TPS on an RTX 6000 Blackwell 96GB through my own coding agent . Opus clearly understands voxel physics. Terrain holds, chunks line up, collision works. The 27B c...

I've been meaning to post about this. The community has been pretty vocal in criticizing "vibe-coded" projects. I used to think the backlash was the real problem, but I've started getting annoyed by a lot of these posts myself — many are just tiny, hyper-specific tools with minimal impact. Still, I think the community and mods could create better spaces for sharing actual ideas and innovations so people can build on each other's work. A monthly mega-thread or "top picks" ro...

From: Vladik on 𝕏: https://x.com/Kostoglodov/status/2071144065857679631 Shaw (spirit/acc) on 𝕏: https://x.com/shawmakesmagic/status/2070918006033817867

sycl : support --split-mode tensor #24152 I'd like to see some numbers if anyone has 2xintel gpus and tries this out

Hey fellow localites, Has anybody used DwarfStar with DeepSeek V4 Flash on 1x DGX Spark yet? What are your thoughts? Based on what I read, with its MoE approach, and its unified memory first then bleed into SSD approach, you can apparently load DS4 Flash and it runs well, with 80B active and full max context. https://github.com/antirez/ds4 There is a video on it here: https://youtu.be/9gHcmhUDJfw Has anybody done this yet? How is the agentic coding quality? Thanks s...

SupraWeather Nano is live! ⛈️ We just released SupraWeather-Nano (preview), a small FT-Transformer model purpose-built to classify weather phenomena from raw tabular meteorological features. https://huggingface.co/SupraLabs/SupraWeather-Nano-Demo https://huggingface.co/SupraLabs Most weather classification setups either bolt a generic model onto tabular data or skip structure entirely. SupraWeather Nano uses a dedicated Feature Tokenizer + Transformer Encoder (FT-Transformer), each i...

When I am talking with Chat GPT or Claude about abstract concepts, I am often surprised by how they seem kind of dumb... like they aren't benefitting from their extra parameters over top open models like Kimi or GLM. In fact they often seem stupider than these open models that I run locally in the way they leap to conclusions and try to oversimplify concepts and produce more "not x but y" like slopisms My guess about why this is: system prompts are lobotomizing the model by trying to...

Are all modern LLMs tuned for chat? Are there any that do bare text completion? I honestly couldn't find any on hugging face.

GitHub: https://github.com/noumena-labs/Sipp

I'm wondering if anyone else has come across this. I've tested the same model on llama.cpp and vLLM with similar settings and quantizations. The performance and concurrency in vLLM are much noticeably better, but sometimes the model feels less reliable. Some things I've noticed: * More mistakes with formatting and tool calls * Forgetting context suddenly * Sometimes acting like messages didn't exist * Lower quality code even with similar parameters I'm not trying to start a compari...

Hey all, So I have a Nvidia DGX Spark and an AMD Strix 395, both have 128GB of unified memory. The Spark has 200Gbit network and the AMD Strix has 5Gbit ethernet (but it has a pcie gen 4x4 slot). Is there any chance I can cluster the 2 together to run a larger model that can fit in the ~256GB (minus OS) unified memory? I've seen that Deepseek v4 Flash can fit on 2x DGX Spark, but maybe I can use my Strix system instead? Any ideas on if this would be possible? If so, how would you go abou...

I got one of these refurbished units last year. I have nothing but good things to say about it. It works great. It has heft to it to keep the GPU secure. And unlike some cheaper Oculink docks, it has redrivers for signal integrity.

With all this talk about Mythos being able to hack into. US government systems, I was wondering if anyone has tried to get root on their own system using a local model?

TL;DR: the recipe's image-build mods aren't actually public – I reconstructed them from the public kernels (with Claude) – and you have to build vLLM at the author's exact pinned ref or the real AWQ weights crash on load. Running now at ~9.4 tok/s on my own 4× GB10. Saw a link on X to CosmicRaisins' GLM-5.2 stack for 4× GB10: vLLM TP=4, MTP speculative decode, ported sparse-MLA Triton kernels (the Hopper-only _flashmla_C path doesn't exist on sm_121), and a data-free 15% expert prune so the AW...

Just found this blog : https://blog.hellas.ai/blog/thunderbolt-ibverbs/ Experimental implementation of RDMA, demonstrated on two Strix Halo Did a quick search and can't really find it posted before ? This could be huge as it shoud work with any USB4 host.

Hi. I am self-hosting Qwen 3.6 27B Q8_K_XL with Llama.cpp on 4x5070ti. (All 4 cards are on single x16 slot bifurcated to 4x4 with risers). I've been testing it on several work repos with Opencode CLI and in like 8/10 situations the output of non-MTP model is far superior to the MTP ones. The prompt is simple `Do a code review of this branch.`. The non MTP produces more findings, with more detailed descriptions, with fix suggestion snippets, everything is better. Usually takes fewer token...

https://preview.redd.it/00o5xtaznf9h1.png?width=696&format=png&auto=webp&s=60a3306ea86a9b0d1f58c435b7dbb0a42761a415 Apple raised the prices across the product line this morning: https://www.reuters.com/world/asia-pacific/apple-raises-prices-macbooks-ipads-memory-costs-skyrocket-2026-06-25/ Beyond the base price, the cost of memory upgrade also doubled. Some stores like bestbuy hasn't updated their prices yet, place your orders when you still can! wondering what this means...

I needed simple local image generation without the usual setup. No virtual environments, no ComfyUI with a complex graph and installation as an exe. So i tried to push the whole thing into the browser and run it on WebGPU. It's a browser extension. You install it, then it loads model, and after that it runs on your own GPU, offline. It use text encoders, UNet, and VAE are ONNX graphs, running on the browser's WebGPU stack. Github : https://github.com/d0grr/generate-ai-images Firef...

Building an inference cluster to run GLM 5.2 locally. Got a Dell quote locked at $8,959.99/unit for 6x RTX PRO 6000 Blackwell Max-Q (300W). List price just jumped to $15,999 yesterday. Quote expires in ~3 hours and I can't swing all 6 right now. I have a second quote for 2 units at the same price that's good until July 3. What would you do — buy the 2 and eat the price increase on the rest later? Try to find someone to split the 6-unit order? Something else? EDIT: I didn't come here for fi...

Including 9B Dense, 31B Dense, 35B MoE, and 397B MoE and reporting sota on different benchmark (let's see if this holds). https://huggingface.co/collections/deepreinforce-ai/ornith-10

I read it with a little bit of effort The tiny model result is insane, theoretically this could make make a 0.5b on-par with a 2/3/4b ish class model in coding with no weights change*. And for large models it could maybe fix let's say 30-50% hallucination problems (educated guesstimate here) Don't expect this to ever come to vLLM or SGLang, but llama.cpp could integrate this easily* like `--top-n-sigma`. *Now there's this one... small... okay big catch: Aside from this being a backtrack...

Hello guys, hoping you're doing fine! I was wondering, for users with 4x-8x 6000 PROs (so between 384 and 768GB VRAM), how are bigger models working for you? I have planned to either jump to 4 or 8 from my actual system, and want to see the experiences with these lately. In theory you can run GLM 5.2 at 4 bits, but not 8 bits right? Same with Kimi 2.7, or DeepSeek V4 Pro. There is a ton of info here https://github.com/local-inference-lab/rtx6kpro/blob/master/benchmarks/results.md , but mi...