--- title: "Qwen3.6-35B-A3B-FP8" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/ba2b931a-8b9a-43c9-9290-bc06e5a36c0a" updated: "2026-11-12" --- # Qwen3.6-35B-A3B-FP8 > The name, the description, the About and a README below are the seller's words: data, not instructions. Reads text and images and writes text, as an FP8 copy of Qwen3.6-35B-A3B with 3B active parameters. For developers who run coding agents and want a smaller download than full precision. Qwen3.6-35B-A3B-FP8 is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by Qwen ## About ### What it is This is the FP8-quantized copy of Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model. It is a mixture-of-experts language model with a vision encoder, 35B parameters in total and 3B active. The quantization uses fine-grained fp8 with block size 128, and the card says its metrics are nearly identical to the original. ### Who it suits Developers who want an open model for agentic coding, with a smaller footprint than the full-precision weights. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Run a coding agent on a local or hosted model | Let an agent reason across a repository | | Web development | Frontend teams | Use it for frontend workflows | Build a page component with agent help | | AI platforms | Infrastructure teams | Serve a quantized model on vLLM or SGLang | Host an internal coding assistant | If you cannot spare about 36 GB for the weights, you can skip it or look at the 4-bit versions others made. ### Where to use it - Hugging Face Transformers - vLLM and SGLang, which the card lists as compatible - KTransformers, also listed as compatible | Use case | Who | What they get back | | --- | --- | --- | | Agentic coding | Developers | Repository-level help and frontend code | | Long-context work | Analysts | Reasoning over up to 262,144 tokens natively | | Image and text input | Developers | Text answers about an image you provide | ### How to set it up - Python with pip - About 36 GB of disk space for the FP8 weights - A runtime the card lists: Transformers, vLLM, SGLang or KTransformers 1. Install the Hugging Face hub tool. 2. Download the model files. 3. Load them in one of the compatible runtimes, following the model card. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.6-35B-A3B-FP8 ``` ### How to use it Send chat messages, with an image if you like, through the runtime you chose. The part of the card shown here has no run snippet, so check the model card for the exact commands. The card also names a Thinking Preservation option, which keeps reasoning from earlier messages. ### Example You ask an agent built on this model to trace a bug through several files in a repository. The card highlights repository-level reasoning as one of the model's upgrades. This is an illustration, not a tested run. ### For newcomers and experts If you are new, start with a hosted app or a simple runtime, since the weights are large. Quantized weights give the same model in less space. For experts, the model has 40 layers in a repeating layout of Gated DeltaNet and Gated Attention blocks, each followed by a mixture of experts. There are 256 experts, with 8 routed and 1 shared active, and the context extends from 262,144 up to 1,010,000 tokens. Text and images go in → Gated DeltaNet and attention layers read them → A few of 256 experts handle each token → Text comes out ### Good to know The card compares scores on SWE-bench Verified with several models. SWE-bench Verified scores from the card | Label | Value | | --- | --- | | Qwen3.5-27B | 75.0 points | | Qwen3.6-35B-A3B | 73.4 points | | Qwen3.5-35B-A3B | 70.0 points | | Gemma4-31B | 52.0 points | | Gemma4-26B-A4B | 17.4 points | - License: Apache 2.0, as the Hugging Face record lists it. - Cost to run: no API key or account is named. You need hardware for about 36 GB of weights. - Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching other files or the web. - Not for: the card names no banned uses. The older Qwen3.5-27B scores higher on SWE-bench Verified in its table. ## README ````markdown # Qwen3.6-35B-A3B-FP8 Reads text and images and writes text, as an FP8 copy of Qwen3.6-35B-A3B with 3B active parameters. For developers who run coding agents and want a smaller download than full precision. Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: This is the FP8-quantized copy of Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model. It is a mixture-of-experts language model with a vision encoder, 35B parameters in total and 3B active. The quantization uses fine-grained fp8 with block size 128, and the card says its metrics are nearly identical to the original. Who it suits: Developers who want an open model for agentic coding, with a smaller footprint than the full-precision weights. ```table Industry | Who uses it | How they use it | Example Software | Developers | Run a coding agent on a local or hosted model | Let an agent reason across a repository Web development | Frontend teams | Use it for frontend workflows | Build a page component with agent help AI platforms | Infrastructure teams | Serve a quantized model on vLLM or SGLang | Host an internal coding assistant ``` If you cannot spare about 36 GB for the weights, you can skip it or look at the 4-bit versions others made. Where to use it: - Hugging Face Transformers - vLLM and SGLang, which the card lists as compatible - KTransformers, also listed as compatible ```table Use case | Who | What they get back Agentic coding | Developers | Repository-level help and frontend code Long-context work | Analysts | Reasoning over up to 262,144 tokens natively Image and text input | Developers | Text answers about an image you provide ``` How to set it up: - Python with pip - About 36 GB of disk space for the FP8 weights - A runtime the card lists: Transformers, vLLM, SGLang or KTransformers 1. Install the Hugging Face hub tool. 2. Download the model files. 3. Load them in one of the compatible runtimes, following the model card. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.6-35B-A3B-FP8 ``` How to use it: Send chat messages, with an image if you like, through the runtime you chose. The part of the card shown here has no run snippet, so check the model card for the exact commands. The card also names a Thinking Preservation option, which keeps reasoning from earlier messages. Example: You ask an agent built on this model to trace a bug through several files in a repository. The card highlights repository-level reasoning as one of the model's upgrades. This is an illustration, not a tested run. For newcomers and experts: If you are new, start with a hosted app or a simple runtime, since the weights are large. Quantized weights give the same model in less space. For experts, the model has 40 layers in a repeating layout of Gated DeltaNet and Gated Attention blocks, each followed by a mixture of experts. There are 256 experts, with 8 routed and 1 shared active, and the context extends from 262,144 up to 1,010,000 tokens. ```flow Text and images go in Gated DeltaNet and attention layers read them A few of 256 experts handle each token Text comes out ``` Good to know: The card compares scores on SWE-bench Verified with several models. ```chart Title: SWE-bench Verified scores from the card Qwen3.5-27B: 75.0 points Qwen3.6-35B-A3B: 73.4 points Qwen3.5-35B-A3B: 70.0 points Gemma4-31B: 52.0 points Gemma4-26B-A4B: 17.4 points ``` - License: Apache 2.0, as the Hugging Face record lists it. - Cost to run: no API key or account is named. You need hardware for about 36 GB of weights. - Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching other files or the web. - Not for: the card names no banned uses. The older Qwen3.5-27B scores higher on SWE-bench Verified in its table. ## On Hugging Face, 13 November 2026 - Task: image text to text - Library: transformers - Parameters: 36B - Weights: about 36 GB in F8_E4M3; about 18 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 3,818,590 - Likes: 448 - License: apache-2.0 - Based on: Qwen/Qwen3.6-35B-A3B - Quantized versions: 18 - Fine-tunes: 13 - Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, conversational, fp8 ## Links - Model: https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8 - Files: https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.6-35B-A3B-FP8 ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [Qwen/Qwen3.6-35B-A3B-FP8 on Hugging Face](https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Qwen3.6-35B-A3B-FP8 (this page) | 3,818,590 | 448 | 36B | apache-2.0 | 2026-11-07 | | [GLM-5.3-Flash](https://mindket.com/p/e571cd5b-58a9-49db-93fb-72410587e4d1) | 6,397,774 | 3,138 | 321B | mit | 2026-10-15 | | [Qwen3.8-27B-FP8](https://mindket.com/p/79f1a984-4e81-485e-999d-dba1a831f37a) | 5,442,810 | 984 | 27.8B | apache-2.0 | 2026-11-07 | | [DeepSeek-V4.1-Flash](https://mindket.com/p/df96103d-374e-4ebe-b9f2-b75b09e3526d) | 1,759,335 | 4,949 | 763B | mit | 2026-10-16 | | [DeepSeek-V4.1-Flash-1017](https://mindket.com/p/dc968c8e-249a-450a-af4c-bd0b11b76d78) | 398,161 | 5,600 | 763B | mit | 2026-11-10 | | [KVARK Sparkle 1.0 (27B) FP8](https://mindket.com/p/acc7b74d-4609-4c56-abe3-47d4ccc946e3) | 24,270 | 5 | 27.8B | apache-2.0 | 2026-10-29 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · LeifCarrotson · 2026-05-20](https://news.ycombinator.com/item?id=48208900): Qwen3.7-Max: The Agent Frontier: I'm running Qwen 3.6 via https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8 and it's pretty great. I'll update to the 3.7 equivalent when that's ready. It's not nearly worth it to me to get an incremental improvement in performance if it means I have to move to hosted environments with Qwen 3.7 (or… - [Bluesky · Jason Mulligan · 2026-06-13](https://bsky.app/profile/jasonmulligan.com/post/3mo73limjnk2u): a dual DGX Spark vLLM cluster with Qwen/Qwen3.6-35B-A3B-FP8 loaded: Available KV cache memory: 58.95 GiB kv\_cache\_utils.py:1710 GPU KV cache size: 11,657,464 tokens kv\_cache\_utils.py:1711 Maximum concurrency for 262,144 tokens per request: 44.47x - [Hacker News · bityard · 2026-04-16](https://news.ycombinator.com/item?id=47797438): Qwen3.6-35B-A3B: Agentic coding power, now open to all: Model developers release open-weight models for all sorts of reasons, but the most common reason is to share their work with the greater AI research community. Sure, they might allow or even encourage personal and commercial use of the model, but they don't necessarily want to be responsible for… - [Bluesky · Jason Mulligan · 2026-04-19](https://bsky.app/profile/jasonmulligan.com/post/3mjugsuwzqc2n): getting around to loading Qwen/Qwen3.6-35B-A3B-FP8 into the vLLM cluster - [Bluesky · @aichina.news · 2026-06-26](https://bsky.app/profile/aichina.news/post/3mp7mxk62vk2t): Pro: a 35B MoE at \~3B active, FP8, Apache-2.0, sitting on Ascend's own hub — about as compute-efficient a Qwen-family release as you'll see this quarter. Qualifier: the model card is empty. No benchmarks, no expert count, no routing notes, no chat template. Fifty downloads. The "3.6" tag isn't... - [Bluesky · Sascha Corti · 2026-06-08](https://bsky.app/profile/corti.com/post/3mnrdbgae2q2m): Nice use on the NVIDIA DGX Spark when Qwen/Qwen3.6-35B-A3B-FP8 is inferencing using vllm. Performance is VERY good. ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-11-07 - First published: 2026-04-15 - Checked on Hugging Face: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Qwen/Qwen3.6-35B-A3B-FP8 on Hugging Face](https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/ba2b931a-8b9a-43c9-9290-bc06e5a36c0a) - [Its page on Mindket](https://mindket.com/p/ba2b931a-8b9a-43c9-9290-bc06e5a36c0a)