GPU guide (GB per dollar, bandwidth)
hope that helps my script uses the GPUs most discussed on the LocalLLaMA, LowEndLocalAI, and LocalLLM subs. At first, I tried to include more, but it became unreadable. Prices were collected by ChatGP
Knowledge catalogue
hope that helps my script uses the GPUs most discussed on the LocalLLaMA, LowEndLocalAI, and LocalLLM subs. At first, I tried to include more, but it became unreadable. Prices were collected by ChatGP
I'm the CTO of a mid-size nonprofit, and I run a local Qwen stack as a coding subagent for that work — partly on principle (a fair amount of what I handle shouldn't leave the building), partly because
Hello LocalLLaMA Fam! I built this project because I, like many of us here, enjoy local AI and have a nice setup, but still find myself reaching for Cloud-based AI due to their convenience. My models
I found a lot of room on the table for these cards so I decided to make a specialized build to squeeze all I could. first The results: qwen 3.8 next Q3_K_XL: 920tk/s pp8192 (2 cards, ram offload), 24/
Ling-3.0-flash-VL inherits the language, reasoning, and long-context capabilities of Ling-3.0-flash, while extending them with native image and video understanding. The model has 124B total parameters
I’m currently running llama.cpp on two RTX 2060 12GB cards, so 24GB total VRAM. With Qwen3.8 27B IQ4_XS at 131k context I’m getting around 45 tok/s, which is actually pretty good for this setup. I fou
I was reading up on the recent controversy around Tristan Buckmaster, Levent Alpöge, OpenAI, and the Navier–Stokes result, and it got me thinking about something broader than this particular dispute.
Privacy have been concern of many of us to have their own hardware to run llms, and here's another reason why: two mathematicians spent a year cracking one of the hardest problems in math and fed ever
I was inspired by Bijan Bowen video - Subway FPS https://youtu.be/6kjXzTVmT58?t=1035 Wondered how far I can push Qwen 3.8 27b so I used a plan made by Fable 5.1 DESIGN.md which has 267 KB! ( 26K of de
Up front: I'm one of the people building the page-perception layer used here. We started by testing small local models. The result turned out to be more interesting than the original test. 12 small mo
Hey guys, After my CPU-only to 96GB VRAM test, I tested Qwen3.8-Flash-Next across llama.cpp, SGLang and FreeToken on the same workstation. This time I wanted to see what changes when you keep the hard
An update to my previous post on running Flash-Next with the expert cache and MTP. I found another useful improvement on the same dual-3090 setup: replacing the CUDA top-k fallback used by my build. O
I don't think anyone posted about this here, but Qwen released a finetuned version of 3.5 4 for driving. The full Bf16 checkpoint is 9B. This is a very interesting development of Chinese AI labs tackl
Gemma 4 12B runs on an RTX PRO 4500 Blackwell. Gemma 4 E2B run on a Jetson Orin NX 16GB; similar performance is expected on a Jetson Orin Nano Super 8GB. Both systems use a reSpeaker Flex 4-mic array
I'm looking for more information to expand my creativity around this. I don't really have a realistic idea of what people actually do with local AI yet, I mostly just want to explore the possibilities
I’ve been thinking about this after using more agentic/local coding models. A lot of the newer models are surprisingly good at continuing on their own. But sometimes that seems like the problem. If a
Quick setup on linux: STEP 0: install/download llama.cpp, Freecad, your favourite gguf model - possibly with multimedia image reading capabilities (I've used Qwen3.8-27B-UD-Q4_K_M and relative mmproj-
Hi all! I just wanna say that I am tired lol. Yes, it's another harness, but I spent a lot of time and effort and have forsaken my hobbies to build the Jenny (like XJ-9) app. Jenny is a free, MIT lice
I wanted a quick calories counter for myself, using LLMs to evaluate the calories from pictures of meals + descriptions. I needed to pick a model so I made a quick benchmark. The setup was: - Nutritio
Please, Oh gentle soul, please beautiful soul, please kindhearted soul, with serious hardware and time to kill, please benchmark Unsloth Qwen3.8-27B-Q8_K_XL against Qwen3.8-Next-Flash-Q4_K_XL and prov
This weekend I posted about the gap closing between frontier models and open source models. Well, now I'm coming with receipts. I've been running local + cloud models against real public github codeba
Model: DeepSeek-V4-Flash-Vision-Exp (local and API when impatient) Time: about one weekend (2 days) of QA and small improvements Full game is here After Qwen3.8-Flash-Next one-shotted a really cool Ca
What can I do to get more from MTP if anything? It am I sitting at the expected range. Thanks /opt/llama.cpp/build/bin/llama-server --host 10.210.44.126 --port 11434 --model /opt/models/Qwen3.6-35B-A3
I got 2x 20GB RTX 3080s + 128GB of DDR4 2666hz RAM (only 4 of 6 channels populated) + a Xeon 6148 I've always been a llama.cpp person and I've been running Unsloth's Q4_K_XL quant of Qwen 3.8 Flash Ne
I moght get shit on for posting this but, I feel like i don't see this being talked enough and it feels like such a waste of a good piece of software. Exl3 is incredible, albeit only if you have NVIDI
The missing control is visible in sudoingX’s Ling-3.0-flash benchmark graphics. The earlier table leaves Ling’s no-speculation baseline as “not measured.” The later code/prose graphic fills it in: abo
The LLM is limited to NPC emulation. Game state, world logic, quests, and the authored story are handled by deterministic game systems rather than the LLM. I built it this way because I wanted the fre
It just seems every local 30b class model is just trying so hard to be the next Qwen that they all just kinda blend into a mass of code focused models. I really like how gemma 4 31b turned out with it
LayerStoRm: GLM-5.3-Flash UD-Q4_K_XL (186 GiB) at 1M context on 2× RTX 5090 + 2× RTX 5080 (96 GB VRAM total) using RAM for the pinned experts. LayerStoRm is a (still experimental) MIT-licensed continu
Back to reading about LLMs as agents on your phone doing GUI tasks! This time I read about MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive and MCP-Augmented Environments a
Hi, I use hermes as a harness, and I am pleased with it, but sometimes Hermes's context size is a bit too much for my system, so I wanted to delegate or straight up use a lighter option for coding tas
DISCLAIMER I ramble a lot and am not good at focusing on documentation and or coherent thoughts. TLDR I did things I made a thing and it works for me. I will try to answer any questions. A few people
TL;DR My TAK quant of Qwen 3.8 27b scored 82.81% on reasoning, comparted with 77.34% for the byte matched Unsloth UD IQ2_S and 83.59% for BF16. Edit: Some of you have tried coding with this reasoning-
How it works. The model being tested is given a server capable of running it's weights and full context. That server is placed in a median priced apartment. The AI is given a bank account with for ren
Long time user of 3.6 27b, switched over to Next Flash since it's a logical step up even from 3.8 27b. It's soooo verbose, i'm talking 13 minutes of thinking time on single turn coding requests at app
Since the Qwen3.5 0.8B model is an interesting one for small specialized fine tunes, I was curious how fast it can run on CPUs. Why CPUs? Mainly because I want to use it as a local dictation cleanup m
My XFX Radeon RX 7900 GRE 16GB Vram GPU struggles with models over 20B size. I added my Radeon RX 480 8GB Vram GPU to the system and ran a few benchmarks using llama.cpp Ubuntu Vulkan build 10453. htt
https://preview.redd.it/3p6234jzk2oh1.png?width=900&format=png&auto=webp&s=b9eab3351d6c6dbd4d5b3fd677a6c40b57f18167 https://huggingface.co/tencent/EVIE-8B 🌟 Highlights SOTA Retrieval Performance: 66.7
I've hit again a point where me as a developer have to take a break from all this slop shit. Im a Developer for 13+ years and i loved it. But i fell for the slop trap. First it started with copilot an
So, I’m currently just short of being able to run the mainline dense SLMs (Qwen 27b/Gemma 31b), running a i9-14900F/RTX4070 Super/32GB DDR5 setup w/ a 700W 80 Gold PSU and a MSI PRO B760-VC WI-FI IV m
I've been tinkering with local LLMs since the beginning of the year when I had an Intel Arc B580 and 32 GB of DDR5. Curiosity got the best of me and I bought the first R9700 about half a year ago, als
TL;DR: switching KV cache to f16 may give a boost in speed if using MTP and ngrams. I have a self-built 'AI mega-cluster' with 2x P40s on a cheap Chinese motherboard and a Xeon CPU (around $1,100 to b
Was considering choosing a 4xRadeon AI Pro R9700 build for a office AI server. Curious to know what results you guys are getting? Mainly considering DeepSeek V4 flash and Qwen 3.8 Flash with offloadin
This comparison was requested by a few people, and certainly we were all eager to see the final results. The comparison had taken 11 days and the GPU was crunching numbers for ~167 hours. We've been c
Hi everybody, I am in a tricky situation. My grandfather + grandmother live in a remote village in asia, and were visiting us in the city. My grandfather has absolutely fallen in love with AI, and it
So after all my work, yeah, Raymond did it better, so I ported his work over, extended it turboX, extended it multiple other models (he had only Qwen models), and benchmarked the crap out of it to mak
While we see for frontier models similar scores among famous coding benchmarks, across: DeepSWE, Terminal-Bench, LiveCodeBench, Code-Arena ELO. Here are in my opinion some next level benchmarks that r
I'm using DS-V4-Flash-Vision with Q8_K_XL quantization locally as my everyday engine, and for some time now I've been doing a lot of comparisons with Qwen3.8-Flash-Next, also with Q8_K_XL quantization
Hello everyone. I felt current LLM benchmark harnesses hand you headline numbers but offer no tooling to see how models actually answered each question (they dump everything to a JSONL or Parquet file
I only use free tiers of these large models to offset compute while my own system runs and for 'different' points of view, since what pops ups suggestions seems to vary a lot sometimes, even when buil
https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF https://huggingface.co/XHToken/Spark-X2.5-1.7B-GGUF from XHToken: We are introducing Spark-X2.5-4B and Spark-X2.5-1.7B, two compact, general-purpose l
Hi, I've been wanting to run my own local LLM for some time and I finally saved enough to get a budget GPU. I can spend 700 at most, and I'm looking for a GPU that can run quantized 30B~ parameter mod
Right, so this is going to be embarrassing but it's presumably something we've all been through at one point or another, and I guess this is my first time resolving something like this in the way that
Hello, I'm a bit obsessed with cache management on local LLMs. For the last few days I've been working on cache management on vLLM with my 2x DGX Spark and DeepSeek v4 Flash 0731. I felt something was
You need Blender 5.x with MCP extension. I use pi so I also need: pi install npm:pi-mcp-adapter and then: $ cat .mcp.json { 'mcpServers': { 'blender': { 'command': '/home/jacek/.local/bin/uvx', 'args'
https://village-sim-one.vercel.app/ - 16GB VRAM RTX 5070 Ti, fully offloaded - Vision on CPU - Windows, not headless - beellama.cpp - latest version with the kvarn performance enhancements making it a
Hi everyone I'm running four rtx 4090, 64GB ram, on a threadripper pro motherboard so all PCIe x16 ports, as a homelab machine for coding. I was migrating from vllm + qwen 3.8 27B (fp8+256k kv cache)
Hugging face has over 300 versions of qwen3.8-27b. I have no idea how to identify the beat model to download and use. Should I sort by “most likes” or “most downloads”? Is there any sort of other guid
I just spent the last 2 hours of my life on a Friday night debugging a strange error in a prod CLI app. EF core was receive a readonlyspan during a Contains query. Normally, this query converted to a
Time for another update! We have been busy and managed to improve the gains substantially (mostly from exploring existing llama cpp PRs and adopting relevant things). Among other things the README.md