Search: “r-localllama”

DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI. It uses an architecture where everything is a plugin, and is powered by Cordis, whose design is described in A Programm

There's no training anywhere in this. I ported Doom's actual rendering algorithm into transformer weights using a compiler I wrote (torchwright) -- every weight computed, none learned. The prompt carr

dots3-note preview is the first open-weight model in the dots3 family. It is a Mixture-of-Experts model with 280B total parameters, 16B activated parameters, and support for a context length of up to

Some people told me that the difference in richness and layout between Glimmer and Qwen wasn't clear to them. This example makes it super clear. I'm aware that comparing Glimmer 30B (a dense model) wi

I managed to get Qwen3.8-2.4T-A95B running locally with llama.cpp on mu PC just for fun, cause why not. I was using the Unsloth Qwen3.8-2.4T-A95B-UD-Q1_0 GGUF quantization. The full GGUF is about 397

Qwen just released their first 3.8 model. The main addition in 3.8 is prompt-steered reasoning effort. You can tell the model how deeply to think by setting reasoning_effort to xhigh, medium, or low.

Ive been experimenting with task-aware GGUF quants for months, taking inspiration from TASA and TAQO but pushing the allocation lower to to the tensor level. The basic idea is to generate a custom ima

Overview ggml_cpu_fp16_to_fp32 leverages hardware F16C intrinsics (AVX-512, AVX2, etc.), faster than the software-only ggml_fp16_to_fp32_row, bringing 17-31% gain in prompt processing rate for a small

🧩 DeepSeek Harness v0.1 is now available in Developer Preview! 🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license. 🔹 Powered by the Co

To my fellow crazies, the few. Those who dared wrestle with llama-70b, mistral-large, goliath, mistral8x22B, DeepSeekV2/3, wept when llama4 behemoth was announced, picked yourself up and are now wrest

Same old prompt, just appended a TIP in the end: 'Write a single HTML file with a full-page canvas and no libraries. Simulate a realistic side-view of a moving car as the main subject. Keep the car vi

Hey everyone, Non-native speaker, writing my post by hand, let me know if I make mistakes (can only learn from it!) Muse Glimmer 30B is so far quite nice, but I haven't found a clear-cut case yet what

Diffusers has a PR with deets: https://github.com/huggingface/diffusers/pull/14456 Minimax is working on this repository right now and put up a bunch of samples: https://github.com/MiniMax-AI/music3-d

Given Gavin Bakers recent interview where he says: > Multiple private companies are planning on spending at least 2x more per GPU for compute as contracts roll-off and some have spoken about this publ

https://preview.redd.it/nqr9nj5028jh1.png?width=1516&format=png&auto=webp&s=47af40a0737429dc65f2e455b87b2db81725cb14 So I wanted to see... is it possible/viable to run this perhaps once in a while som

Stumbled across this new vision model, SenseNova-Vision. It's a 7B MoT model, Apache 2.0 license, which is cool. The main idea is it treats pretty much all computer vision stuff as just one generation

If you haven't checked the paper: https://arxiv.org/abs/2608.09867 TLDR: the authors show that you can swap out the 'encrypted' reasoning of the biggest model, like Opus, Sol, and put them into weaker

I've been googling 'tar extract gz' for about ten years. and I finally did something about it. It started out as a research project and I ended up with a Fine-tuned Qwen2.5-Coder-1.5B on 125k natural-

They highlighted 3 things on countdown page: VLM, Agentic Improvements, and Think mode. What improvements do you expect? Reply here! Personally, I want to meet a sage who has attained enlightenment. T

In OCP APAC 2026, Tai AMD SVP of compute and enterprise AI said agents don't cut GPU demand but they just pile on a whole extra layer of orchestration, retrieval, and tool-calling work that runs on CP

North Micro Vision Instruct is a 2.4B-parameter open-weight vision-language model with native-resolution image support, released under the Apache 2.0 license. It is designed as a compact foundation fo

Since lot's of people were sceptical or whatever, heres how to uncensor / jailbreak V4 flash and proof. No it is not lead on whatever, first prompt, first try, every time. Put this in System message:

Link to the article: KV Cache Quantization on Gemma 4 31B: Non-QAT vs QAT KLD benchmarks with BeeLlama.cpp v0.4.3, fork of llama.cpp with more KV cache quantization options, comparing Gemma Q4_0 non-Q

Yesteday a paper showed a gap that allows to see 100% of the reasoning tokens form ALL Claude and GPT models Stealing Reasoning Traces from Proprietary LLM APIs. check it out, they have published lots

Hey guys, Just finished benchmarking DeepSeek V4 Flash 284B + DSpark on a single RTX PRO 6000 96GB. Short version: DSpark: ~15–17% faster generation on my coding workload On this setup, the DSpark dra

Sometimes you have tasks that are outside of your expertise and the idea is this workflow automation could be leveraged to manage to have local AI figure it out using research from his workflow gather

I’m not super knowledgeable of the entire AI industry, but as we see this industry grow and the Cold War that is happening between the US and China on AI development, I can’t help but notice what US c

Liquid AI put out LFM2.5-VL-3B today, which is a 3.1B vision model that weighs roughly 2GB and fits well on a phone Benchmarks are benchmarks so I tried something sillier. Took a photo of a little Ste

LFM2.5-VL-3B is a multimodal variant of LFM2.5, a family of hybrid models designed for on-device deployment. It builds on LFM2-VL-3B with further mid- and post-training. LFM2.5-VL-3B can process both

Hey Folks, I've been making quants for a while - recently I took a short break to get into hardcore research (submitted my first EMNLP paper during it!). Along the way, I built up a little arsenal of

i was hoping they would release both today but the big one just dropped now , and it says its been listed since 5 hours ago on huggingface. I guess the real date for 27b is https://modelscope.cn/model

Muse Glimmer 30B feels significantly more precise and reliable, it almost never drops the ball or breaks rules. However, its designs lack creative depth and richness. Qwen3.6 35B, on the other hand, i

One of the reasons I got into local LLMs was the possibility of getting answers using my own documents and books (a few hundreds) instead of having to search through them manually. However since I'm n

I also have 64 gb ddr4 ryzen 5600 Using llama.cpp Ubuntu distro Settings are as follows --n-gpu-layers 999 --n-cpu-moe 36 --no-mmap -ctk q8_0 -ctv q8_0 -fa 1 -c 9000 So rebuilt my llama.cpp build to r

Ran the model with quants (Q5) and MTP by bartowski with llama.cpp server. It takes ~24GB ram running on M5 Pro with 48GB at about 65t/s. On some tasks it was quite the overthinker. Overall, the quali

I'm trying to figure out GPU Kernel optimization on older hardware like SM80(ampere) . Is there tools you guys use? Or frameworks? Im waiting for this framework https://www.reddit.com/r/LocalLLaMA/com

Warning: Kinda long post. If you don't like reading, please skip for your own sanity. Also, I've got nothing to sell, just a tinkerer, so I just want to share ideas and learn from you guys too. When I

A few things right off the bat: it reasons very efficiently. Like Grok 4.5 levels of efficient thinking it quantizes very well. My first few tests with iq3_xxs were better than Qwen/Gemma behaved at t

Hello everyone! ​I like dumb things. I like working with weak computers and microcontrollers. I like the simplicity and low electricity usage. Simply put, the efficiency of a 'dumb' PC. ​The first tim

Seems like we're limited to qwen finetuned MoEs for now. Looking at the current landscape - focus seems to be on dense models (muse glimmer 30b, qwen 3.8 27b) for smaller setups. Is upgrading to 24GB

I just heard of this model. Seems to be a competitor to DeepSeek V4 Flash. About the same size and active parameters. Anyone tested it compared to V4 Flash? Link: https://huggingface.co/upstage/Solar-

Hi! I'm really obsessed with LLM hallucinations for the last 6 days 😭 I started by designing system prompts to attack hallucinations but failed, obviously. Now I tried reading logprobs and... I think

Spent two nights getting deepseek-ai/DeepSeek-V4-Flash-0731 (284B MoE, 13B active, native FP4/FP8, 1M context) running production-grade on two DGX Sparks connected by one QSFP DAC cable. Everything —

Been benchmarking DSv4 Flash 0731 on a Flow Z13 (Ryzen AI MAX+ 395, Radeon 8060S / gfx1151, 128GB LPDDR5X) for the past week. Figured I'd share what actually works and what doesn't — there are a lot o

I'm very happy with some Linux admin tasks I'm throwing at a locally running DeepSeek. My request was simple, check why 'samples' folder is taking more and more space on one of the machines on my LAN,

Everything started with the sudden death of my old ASRock J1900. While looking for the perfect ITX replacement, I stumbled upon the Chinese CW-NAS-ADLN-K motherboard, which looked perfect on paper: In