model
huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4 ↗
9864413 downloads507 likestext-generationModel Optimizer
Model Overview Description: The NVIDIA Qwen3.6-35B-A3B-NVFP4 model is the quantized version of Alibaba's Qwen3.6-35B-A3B model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA Qwen3.6-35B-A3B-NVFP4 model is quantized with Model Optimizer. This model is ready for commercial/non-commercial use. Third-Party Community Consideration This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party’s requirements for this application and use case; see link to Non-NVIDIA (Qwen3.6-35B-A3B) Model Card from Alibaba. References NVIDIA Model Optimizer: https://github.com/NVIDIA/Model-Optimizer License/Terms of Use: Apache license 2.0 Deployment Geography: Global Use Case: Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Release Date: Hugging Face on 05/28/2026 via https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4 Model Architecture: Architecture Type: Transformers Network Architecture: Mixture-of-Experts (MoE) with Hybrid Attention Number of Model Parameters: 35B in total and 3B activated Input: Input Type(s): Text, Image, Video Input Format(s): String, Red, Green, Blue (RGB), Video (MP4/WebM) Input Parameters: One-Dimensional (1D), Two-Dimensional …
UD 3.0 seems to be a massive improvement over UD 2.0 Some of us still want to run the older Qwen models but would benefit from UD 3.0 UD 2.0 vs 3.0 is like the difference between a full quant. So Q3 UD 3.0 is similar to Q4 UD 2.0.
↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6↯ Qwen 3.6qwen
Does anyone know of a decent reverse engineering harness/workspace setup for local models that I can just point the model at and have it go to work until it's reversed most if not all of the functions in a binary, even if it takes days? Of…
Hi everyone! I’m an intern at Icosa, a startup focused on making local AI accessible.
MultiNet 2.0 Preview: Interactive 2D Mazes Frontier Reasoning Agents fail on Interactive 2D Mazes We put 3 highly capable models in 50 simple 2-dimensional mazes each. Claude Opus 4.8, Kimi K2.6, and Qwen 3.6-27B together solved just 6 out…
So here's the thing, almost everyone use NVIDIA to run their LLMs, we also do the same, a lot of people we've met use like RTX PRO 6000 or even H100, B300 It seems like everyone eyes is looking at NVIDIA. However we do the math that the ra…
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive Q4_K_P on llama.cpp, -ncmoe offload cause it doesnt fit in vram outright, 10 threads, q8_0 kv both sides. nothing weird about the setup far as i can tell.
why Qwen 3.6 35b a3b q4-k-m is slower on 7900xtx (20t\s 100% GPU Load) than on 3060ti (37t\s and GPU 50% Load) on the same settings? Linux llama.cpp vulkan 1.5Gb VRAM is empty.
Hi, Back in late march I begun playing around with Qwen 3.6 27b and found like everyone else that it's notoriously good at tool calls, where every model I tried before just derailed after a few turns it kept going and felt quite reliable o…
JetBrains has announced Junie Local, a way to run local models on your machine without having to deal with the manual complexity that comes with installing and running local models. Junie, if you don't know, is an LLM-agnostic AI coding ag…
How We Optimized the Qwen 3.6 Model for Our Junie Agent A while ago, we launched a long-term project to enable users to run Junie entirely locally, with local inference, across a wide variety of hardware setups. After much anticipation, we…
I used to use Qwen3.6-35B-A3B with llama.cpp and connecting it to the VSCodium extension called "Continue." My computer is running a Intel(R) Core(TM) Ultra 7 265K (3.90 GHz) with 128 GB of DDR5 RAM and an Nvidia Geforce RTX 5090 that has…
I have been benchmarking local LLMs on a Mac M4 Pro 24 GB RAM using LM Studio. I've tested mostly with 4-bit quantization, both MLX and GGUF, from 4b to 35b models, with speeds of 3 to 40 tokens/second.
For moe models like Qwen3.6 35B A3B, some setups keep experts on the cpu in order to have the active weights in the gpu. But I have a strix halo + r9700, so I wondered what if I kept the experts on the strix halo and active weights on the…
The caveat: this was a quick benchmark for concurrency, but this holds at depth even more for the CUDA card vs AMD. In addition, MTP is on for strix halo, and off for the CMP170.
When I have a normal conversation with my Qwen3.6 35B A3B I get ~27t/s. I noticed that when I use a harness sometimes it can be slow as 15t/s (tested OpenCode and Maki), with way lower context than my normal chats.
Look at me: I am the frontier Lab now Huihui-Qwen3.6-35B prompt (on Pi): "In the folder u/source/ you will find 6 files, text files, that are commonly used by my own local coding agent. your task is to take each file and create a new versi…
Just make sure to leave MTP off. It's a decent upgrade from Qwen3.6 27B, and better than KAT Coder V2.5 Dev (which also requires their custom vLLM).
I'm finally upgrading from my 32GB unified memory + 780M setup to a single R9700 via eGPU Assuming all goes well I could run a smaller faster model in main memory. What exactly do you delegate to a smaller model?
We tested Qwen3.6 27B quantizations by Unsloth on Hugging Face, with pelicans on bikes, gears, Terminal-Bench 2.1, and AIME-120.
Benchmarking Qwen 3.6 35B MoE (3B active) on an RTX 3090 I mentioned I'd got a second RTX 3090 on a group chat, and a friend said: I know this is not really your thing... but let me know how quickly it runs Qwen 3.6 35bn MoE.
could not extract summary