Knowledge catalogue
Hello! I was looking into quantizing models and i saw how Hy4 was shrunk from 1.5 TB to 200GB with high retention in benchmarks (98% i think). I was wondering if: a) it would be worth it to attempt th
Hey guys, I tested Qwen3.8 Flash with llama.cpp from CPU-only to the full 96GB of my RTX PRO 6000. Short version: CPU-only reached 8.34 tok/s at a 2K prompt Full 96GB reached 109.07 tok/s At 245K cont
## Qwen3.8-Flash-Next-NVFP4 (inferact) vs Qwen3.8-27B-FP8 (qwen) Slammed with work and no time to pretty this up. Qwen wrote most of this but I checked the data. All tests done on the same rig, same p
Mac Studio, 96GB unified memory (M3 Ultra). I want the largest usable Qwen3.8-Flash-Next setup, and I'd rather not burn 100GB of bandwidth on the wrong download. Here's my math. Please tell me which p
Stumbled around a post about optimizing new Qwen up to 178t/s with a patched version of sglang : https://github.com/jpezzulli/sglang-rtxpro6000 I managed to reproduce results (kudos to jpezzulli, whom
TL;DR: Ninfer/DS4 but for RDNA4 Highly custom kernels built for RDNA4, applied to vLLM-Radiance to greatly improve Qwen3.8 Flash Next speeds. This is mostly for dual R9700s with preferably 48GB of RAM
TL;DR - yeah it works, no wild skills or patching required. 37-50 tg/s when generating text, spiking over 60tg/s when writing code and mtp hits hard. Your results can be better if you have a better mo
Interesting new paper from Alexia Jolicoeur-Martineau (of Tiny Recursive Model fame) and collaborators. They seem to be able to replace quadratic attention with sliding window attention + attention si
SlopTV: a YouTube live stream where the chat writes the programming. You type 'capybara dj underwater rave', an LLM inflates it into a 400-word structured video prompt, one of my 5090s renders 15 seco
https://huggingface.co/Nanbeige/Nanbeige4.2-3B-DSpark 4b model for the gpu poor that I think is stronger than qwen 3.5 9b now faster! I was getting roughly 35 t/s AR which was this models downside, it
Sori-1B is a 1B-parameter audio-language model built by a single SNU researcher whose core claim to fame is that its decoder is trained entirely from scratch on audio-paired text — no text-only pretra
A hallucinated event, Crono awakens in his modest bedroom of 2095... I am using local models since 2025 January. My daily driver is qwen 3.6 35B A3B, which works pretty well for coding tasks, but I al
With b10726, the default --lazy-mode change keeps the 51B-parameter PLE n-gram embedding table of Qwen 3.8 Flash Next on disk: it is mmap'd and its rows are read on demand during inference, even with
I used Claude Code to help write this pipeline - Gemma4 to turn an idea into a short story, then IndexTTS 2.5 running on an NVIDIA 3080 to use voices from LibreVox and VTCK voices to pin as character
Hi All, I am currently contemplating an upgrade from my 5060ti 16gb. I am getting ~40t/s with 130k context on Qwen 3.8 IQ3_S HF quant. I am running llama.cpp on linux. Objective is to increase context
So, ideally for this thread we exclude the ones that everyone on here is already well aware of and discussing on here a lot, like N-gram, quantization improvements, MTP, D-flash, and D-spark, since th
When using two GPU with llama.cpp with --split-mode layer you usually provide a --tensor-split value to fit the proportion of the allocation on each GPU, say you have 16GB + 12GB something like --tens
I'm running Muse Glimmer 30B EXL3-SC 3.00bpw H4, fully resident on my 12GB VRAM GPU at 100K context with Q8_O KV cache. It's a joy to use a dense 30B model at this size and still get 30 tok/s on a VRA
(not written by Claude, all errors and crappy text are result of too little coffee on a Sunday morning ;) Our home server is a 2018 Thinkstation P520, bought for about 600eur in 2023. It's been upgrad
Hey guys I have been working on a open source LLM development repo. Anyone can train any amount of parameter against any arch (GQA, MLA, Dense...) When I tried to implement MLA I read the technical re
I’ve been experimenting with MiniMax H3 and finally have video generation working in TensorSharp. TensorSharp started primarily as a local GGUF/LLM inference engine, so getting a video-generation pipe
Since I spent the time to figure it out and it is not like it will make me any money ever. I think I'd share with you all what I managed to cobble together. First and foremost here the relevant bits a
The intelligence in my context is an aggregate index, I called the Agentic Coding Index, across most relevant agentic coding benchmarks: SWE-bench Pro, DeepSWE v1.1, Terminal-Bench (v4, v3, v2.1), Cod
I built SpeakoFlow Mini, an Apache-2.0 fine-tune of Qwen3.5-0.8B for dictation cleanup. It is not a chat model or a general rewriter. It takes speech-to-text output, applies corrections the speaker ac
I see this kind of approach quite often with local AI models. For example, when optimizing models for an RTX 3090, people sometimes convert them from FP8 to INT8 so they can run efficiently on Ampere
I am doing a lot of translation work with different languages, and German is just exceptionally well done by Qwen 3.8 27B. It is lengths ahead of GPT-5.6 and even Fable 5. With the frontier models I o
I used the quantized n-gram to INT4, it's 32 GB, memory-mapped from disk. I confirmed that it works great on 150-160k context, and i was watching all the time my VRAM usage while doing single thread l
After a couple of days of faffing around, here's my current config in case it helps anyone out. Some of this is also valuable for a single Spark. AI slop below for you to read and paste into your own
Running on a M1 Max 64 GB: - SSD streaming for tensors - SSD streaming for engrams - SSD streaming for MTP How is it possible? * Custom Q4 quant: benchmarked all metal tensors then picked and spliced
If you own 4xR9700 and were waiting for the model to make them shine, then I have some good news for you! It's running at 80-120 tokens/second for generation and 12k token/second prefill for a single
Setup: MacBook Pro M5 Max, 128 GB unified, macOS 26.5.2 · llama.cpp b10686 (Metal, 12 threads, batch 2048, flash-attn, kv-unified, ngram-mod spec decode) · Qwen3.8-Flash-Next UD-Q2_K_XL (Unsloth), 78.
Been working really hard for the past month to bring to the community all these models, the hardest was for sure LongCat-Flash-Lite-Sparse who required TONS of work, first I needed to have Heretic sup
I find both Qwen 3.8 27b and Qwen 3.8 Flash Next difficult to read. Here's some examples of what I mean: **Model-visible tool set per turn** (assembled by the host at provider-request time): persona t
We used HFlow to evaluate the latest open weights VLMs for processing egocentric data. This was based on Build AI's Egocentric-10k evaluation, which used Gemini 2.5 Flash to measure hand visibility an
I have been using the unsloth/Qwen3.8-27B-NVFP4 with 157k ctx on vllm currently. But recently I have been seeing post about ninfer lately and how it is like the best way to run qwen 3.8 on RTX 5090 32
Hey all, You’ve probably noticed that using LLMs and AI agents for pentesting has become pretty common lately. The problem is there isn’t really a good way to figure out which model is actually best s
I got a 50% performance boost (20 t/s -> 30 t/s) in llama.cpp for MoE models that don’t fit entirely in VRAM—in my case, Qwen 3.8 Flash Next. The idea is simple: instead of offloading entire layers to
You guys remember sesame AI? With their really good AI voice model? Obviously ChatGPT has their voice model that’s also really good. Is there any smaller local variant that runs on like consumer grade
specs hardware: M4 Max 128GB Studio inference engine: llama.cpp (qwen4exp branch) judge: claude-opus-4-6 AtomicChat/Qwen3.8-Flash-Next-GGUF Qwen3.8-Flash-Next is a great model I benched in my previous
This has been a big 2 weeks for local models with 3.8 27B and Next coming out, a lot of 'frontier labs are done' comments which, in general, could certainly be true. Could be trillions in economic val
Hey all, I stumbled onto something funny and figured I'd share before it gets buried. I'd been running the Ornith 1.5 35B A3B GGUF (Q4_K_M) for a while, then recently a new revision showed up in my HF
In mid-August, Ramp published spending data collected from 70,000 U.S. companies: Fable 5 ,the most powerful and expensive model in Anthropic’s lineup, accounts for just 11% of what those businesses s
I have a 2 DGX Spark setup recently and I have been happily running Deepseek V4 Flash 0731. Since the release of GLM5.3 Flash and Qwen 3.8 Flash Next this week, a lot of folks are still waiting to see
Hardware: rtx 5080 16 gb vram; 64 gb ram ddr5 6000hz; ssd with unlimited memory; ryzen 7 9800 x3d. OS: Windows 11 Software: I’d prefer llama.cpp, but it’s not a strict requirement; I’ll use whatever y
People who moved from older version to this new one, how much improvements do you see on inference? Ex: llama.cpp performance? This new version comes with Linux Kernel 7.0. Hope AMD cards gonna enjoy
Folks! We're just 50 PRs away from more faster inference. Hopefully by end of year. Experts!, please chip in there. List of Open/Ongoing PRs(and also Discussions) related to CPU/RAM/Disk/Hybrid: [Disc
TL;DR: Every public low-bit GGUF of this model is secretly ~4.70 bpw. Shim the rows to 256 and it becomes a real 3.07 bpw / 11.77 GiB file that runs 262K context on 16GB. Needs patched llama.cpp — not
Hey! I forked NInfer (a from-scratch C++20/CUDA inference engine for Qwen models) and added two things: tensor-parallel across two GPUs, and YaRN ×4 rope scaling. Together they let Qwen3.8-27B NVFP4 r
I wanted to share my successful setup for running a Qwen 3.8 27B model with a massive context window on a consumer 16GB GPU (RTX 4070 Ti SUPER). The goal was to fit everything into VRAM without sacrif
I ran a few benchmarks on my MacBook Pro M5 Max with oMLX: Running Qwen3.8-27B with thinking off has a big effect on output quality. Running it with thinking on xhigh burns 5.5x the tokens and runs 6x
Below are the benchmark results for running Qwen3.8-Flash-Next on Strix Halo using the Vulkan backend of llama.cpp, combined with MTP model. Hardware Item Details CPU AMD Ryzen AI MAX+ 395 (16C/32T) G
~20–30GB smaller than Unsloth/AesSedai Q4 at similar PPL. After several days of testing I released a set of mainline-compatible imatrix quants for Qwen3.8-Flash-Next. Goal: same quality band as the po
this model is really something incredible, saved us like 600 dollars. my fiance is always breaking her electronics and then getting me to fix them. The other day she brought me her phone and it was st
I saw someone here posted about getting a Tenstorrent QuietBox 2, and I wanted to look into it the hardware. It's very difficult to find any benchmarks, but I managed to find some from an employee. Th
Announcement: https://www.tbench.ai/news/terminal-bench-4-0 Leaderboard: https://www.tbench.ai/ Imo the best aspect in their announcement is their focus on rapidly iterating on TerminalBench to keep t
The official benchmark card for Ling-3.0-flash-Fin is a useful reminder that the unit being tested is rarely just “the model.” The release says most runs used temperature 1, top_p 0.95 and the highest
hi i make tool for this called llmog it's purpose to make llms free to - auto annotation datasets - reclassification existing yolo datasets running totally local using llama cpp or vllm or use externa
This base model is only 43m parameters trained on 3m arXiv abstracts. We plan to continue pre-training and post training. If you create fine-tuning datasets or if you know of any datasets that can hel
Quote: ... developing a new A.I. model intended to be as powerful as Anthropic’s cutting-edge models. ... In July, while developing Watermelon, Meta paused and later resumed a stage of A.I. developmen
Model: unsloth/GLM-5.3-Flash-GGUF UD-Q2_K_XL (4 shards, 101 GiB) + mmproj-BF16. Reference: llama.cpp PR #27754 (glm5next), CUDA 12.8, SM 120. Throughput Both engines back to back in one session, flash