Instructions to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF",
	filename="Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled.Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Use Docker

docker model run hf.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Ollama
How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Ollama:
```
ollama run hf.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
```

Unsloth Studio

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF to start chatting

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Docker Model Runner:
```
docker model run hf.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M
```

Lemonade

How to use hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF-Q4_K_M

List all available models

lemonade list

12gb vram

by revster1337 - opened Apr 21

Discussion

revster1337

Apr 21

Did anyone tried this with 12gb vram

xpg10

Apr 21

StanBaus

Apr 22

Tested this on my Mechrevo Jiaolong (Ryzen 9-9955HX, RTX 5070 Ti 12GB VRAM, 32GB RAM). Running the Q8_0 version via Ollama, and it works surprisingly well! While the model is larger than 12GB, the split between VRAM and system RAM is handled smoothly. Speeds are decent for daily coding and complex logic tasks. Definitely usable on 12GB VRAM if you have enough system memory.

revster1337

Apr 22

Tested this on my Mechrevo Jiaolong (Ryzen 9-9955HX, RTX 5070 Ti 12GB VRAM, 32GB RAM). Running the Q8_0 version via Ollama, and it works surprisingly well! While the model is larger than 12GB, the split between VRAM and system RAM is handled smoothly. Speeds are decent for daily coding and complex logic tasks. Definitely usable on 12GB VRAM if you have enough system memory.

How many tokens/s ?

jin-xonik

Apr 22

•

edited Apr 22

Yes, RTX 3080 12GB + RAM 32GB DDR4 + i7-12700F, LM Studio.
It writes code with compilation errors, doesn't fix it completely, and then just breaks in the middle of the code. The context is not over yet (32K tokens), Context Overflow = Rolling Windows, Temp = 0.1, Top K = 20, Top P = Off, 10 CPU threads. I don't know how to get it to finish fixing the code (do you know?) 🤷‍♂️
Speed is about 10-20 t/sec depending on quantification (Q4-Q6) and duration of communication.
P.S. Perhaps I'm giving not so trivial request (write a class to generate TOTP codes without external dependencies), but Claude handles it easily.

Request (Russian)

Напиши класс на C++ для создания TOTP-кодов по ключу, time_point любого типа или по текущему времени (т.е. 2 перегрузки, и чтобы при этом было корректное преобразование времени в Unix time) и по длине кода (по умолчанию 6 цифр). Без внешних зависимостей и POSIX. Сделай пример, который выводит кол-во оставшегося времени и 2 TOTP-кода: для текущего времени utc_clock и system_clock. Стиль именования типов, функций и переменных используй привычный для C++, не camelCase. Перепроверь код на ошибки и сразу исправь, если они есть.

jin-xonik

Apr 22

•

edited Apr 22

This model performed the task almost at the level of large cloud models with hundreds of billions of parameters.: https://huggingface.co/Qwen/Qwen3.6-35B-A3B (Q6_K).
With disabled thinking!

revster1337

Apr 22

This model performed the task almost at the level of large cloud models with hundreds of billions of parameters.: https://huggingface.co/Qwen/Qwen3.6-35B-A3B (Q6_K).
With disabled thinking!

What iq rig and config you used

jin-xonik

Apr 23

•

edited Apr 23

What iq rig and config you used

LM Studio
64K tokens, 10 CPU threads, Thinking Off, everything else is by default.
Temp = 0.2, Context Overflow = Rolling Window, Top P = 0.95, Min P = 0.05, Repeat Penalty = 1.05 (but all this is not particularly important AFAIK, except for the temperature).
Speed = 15.6 t/s (1st answer), 12.03 (2nd), 10.13 (3rd, final).

RS14Z

Apr 25

4070ti 12gb, 64gb ddr4, i7-13700f (16 cores).
45 t/s, 100k context
Q4 quant
--temp 0.6 --top-k 20 --top-p 0.95 --repeat-penalty 1.0 --presence-penalty 0.0 --min-p 0.0 --host 127.0.0.1 --port 8000 --threads 12 --threads-batch 16 --parallel 1 --fit on --fit-ctx 92160 --fit-target 512 -n 8192 -b 2048 -ub 512 --mlock --jinja --chat-template-file ./chat_template.jinja --chat-template-kwargs '{"preserve_thinking": true}' --flash-attn on -ctk q8_0 -ctv q8_0

revster1337

Apr 25

4070ti 12gb, 64gb ddr4, i7-13700f (16 cores).
45 t/s, 100k context
Q4 quant
--temp 0.6 --top-k 20 --top-p 0.95 --repeat-penalty 1.0 --presence-penalty 0.0 --min-p 0.0 --host 127.0.0.1 --port 8000 --threads 12 --threads-batch 16 --parallel 1 --fit on --fit-ctx 92160 --fit-target 512 -n 8192 -b 2048 -ub 512 --mlock --jinja --chat-template-file ./chat_template.jinja --chat-template-kwargs '{"preserve_thinking": true}' --flash-attn on -ctk q8_0 -ctv q8_0

Did you tried to compare with 27B model ?

jin-xonik

Apr 25

Did you tried to compare with 27B model ?

Today I added a system prompt for 35B A3B and tried again. It generated a fully working code on the first try! Even Claude Haiku 4.5 succeeded only on the second attempt, leap seconds were not taken into account in the first iteration (the main problem of this task). The speed is 13.1 t/s.

27B (even Q4) is significantly slower because it's not A3B — 3.57 t/s. And the code doesn't even compile (the system prompt was the same).

revster1337

Apr 25

So i should go for the A3B with 12 VRAM

RS14Z

Apr 25

Did you tried to compare with 27B model ?

No. But I think I’ll get something around 15 T/s. I really like 35b-a3b - the speed is excellent. And when connected to the web (search, fetch, ground), the model performs just as well as Sonnet or GPT 5.2. It’s also very important to use “thinking” mode on these models, because without it they’re pretty dumb. In my llama.cpp config, I consider the most important settings to be --fit-on --fit-ctx 92160 --fit-target 512 and -ctk q8_0 -ctv q8_0. Fit-on distributes available resources very well, there’s no need to fiddle with the config to match your specific hardware. With fit-on, it’s important to use fit-ctx for the context, not just --ctx-size. And --fit-target specifies how much VRAM to leave free.

BingoBird

May 2

Did you tried to compare with 27B model ?

No. But I think I’ll get something around 15 T/s.

27b dense with most of your weights on 64gb ddr4 will give about 3 t/s initially dropping down to 1.5 at high context -- not 15.

At least that's my guess. Give it a spin and let us know :)

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment