Qwen 3.6 35B-A3B REAP-48 v2 GGUF

Qwen 3.6 35B-A3B REAP-48 v2 GGUF

Runs a pruned and quantized Qwen3.6-35B-A3B in an 8.78 GiB GGUF file for llama.cpp, closer to the unquantized weights than the first release. For people with limited memory.

Made by crucible-labs 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 a GGUF build of Qwen3.6-35B-A3B from Crucible. It prunes 48% of the routed experts with REAP and then quantizes the rest with bit sizes chosen by measuring each tensor, not by hand. This is the second release. Its file is the same size as the first, 8.78 GiB, but it sits measurably closer to the f16 weights.

Who it suits: It suits people who run a large model locally with llama.cpp on limited memory and want a small file that stays close to the original.

Industry | Who uses it | How they use it | Example
Software development | Developers with a local coding assistant | Run a code model on a 16 GB machine | The card measured 29.24 tokens per second of decoding on an M4 Mac Mini
Local AI hobby | Mac and AMD GPU owners | Pick a small file with less quality loss | The v2 file is closer to f16 than v1 at the same size
Research | Quantization researchers | Study measured bit allocation against a hand-written one | The card compares v1 and v2 by KL-divergence

You can skip it if you use CUDA, Vulkan or a CPU-only build and value speed over fidelity, because the card has not timed the new quant types there.

Where to use it:

• llama.cpp, with the GGUF file
• Windows, macOS or Linux
• Apple Silicon with Metal, where the card timed it on a 16 GB M4 Mac Mini
• AMD GPUs with ROCm, where the card measured accuracy

Use case | Who | What they get back
Local coding help | Developers | A code-capable model that fits in a small file
Tool-calling tests | Agent builders | A model the card scored on a function-calling benchmark
Quality-per-size comparison | Researchers | Measured KL-divergence figures for two releases

How to set it up: You need llama.cpp and about 9 GiB of free space for the file.

• A llama.cpp build for your platform
• Python with the Hugging Face tools, to download the file
• A machine with enough memory for an 8.78 GiB model, such as the 16 GB Mac the card used

1. Install the Hugging Face tools with Python.
2. Download the repository files with the second command below.
3. Point llama.cpp at the GGUF file.
4. If speed matters on CUDA, Vulkan or CPU, benchmark before you switch from v1.

pip install -U huggingface_hub
hf download crucible-labs/Qwen3.6-35B-A3B-REAP-48-v2-GGUF

How to use it: The card gives no run command. Load the GGUF in llama.cpp and prompt it as you would any chat model. The card used llama-bench to measure prefill and decode speed.

Tool | What it does
llama.cpp | Loads the GGUF file and runs the model
llama-bench | Measures prefill and decode speed, as in the card's tests
hf download | Fetches the repository files

Example: You load the GGUF on a 16 GB Mac Mini and start prompting. On that machine the card measured decoding at 29.24 tokens per second for v2, against 26.71 for the first release.

For newcomers and experts: Think of it as a big model that has been trimmed and compressed to fit a small computer, with the loss measured. A first small try is to load the file and chat with it as usual.

Two steps matter most. Calibration now runs through the model's own chat template, where v1 used flattened text. Bits are then reallocated by measurement, because the shared expert turned out to be much more sensitive than the routed experts for each parameter. v2 also adds i-quants, IQ3_S and IQ4_XS among them, while v1 stuck to k-quants. Token generation is limited by memory bandwidth. Each token reads the dense path in full but touches only 8 of 133 experts, so trimming bits on that dense path saves the most time per token.

Prune 48% of the routed experts with REAP
Calibrate with the model's chat template
Measure how sensitive each tensor is
Assign bit sizes per tensor and layer
Write one 8.78 GiB GGUF file

Good to know: The card compares v2 with v1 against the f16 weights, at the same file size.

Title: Closeness to f16, v1 against v2
Metric | v1 | v2 | Change
Mean KL-divergence | 0.034093 | 0.019609 | 42.5% lower
Median KL-divergence | 0.002489 | 0.001186 | 52.4% lower
99th-percentile KL-divergence | 0.536331 | 0.327945 | 38.9% lower
Excess perplexity over f16 | 0.037205 | 0.022679 | 39.0% lower
Top-1 agreement with f16 | 94.678% | 96.199% | 1.52 points higher

Title: Speed on an M4 Mac Mini, 16 GB, Metal
Test | v1 | v2
Prefill, tokens per second | 411.29 | 417.18
Decode, tokens per second | 26.71 | 29.24

Title: Accuracy of v2 in the card's own test setup
HumanEval+: 90.9%
MBPP+: 75.7%
BFCL-simple: 93.2%

• Untested backends: the card did not time the i-quants on CUDA, Vulkan or CPU-only builds, so benchmark first if speed matters
• Accuracy runs: temperature 0, one attempt per problem, measured on ROCm, and deliberately not compared with v1
• Size: 19.2B parameters, about 9.6 GB at 4-bit, which is an estimate
• Language: English
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the prompts you run locally
• Not for: non-English work, since the card lists English only

On Hugging Face, 13 November 2026

• Task: text generation
• Library: llama.cpp
• Parameters: 19.2B
• Weights: about 9.6 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 2,427
• Likes: 34
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.6-35B-A3B
• Tags: llama.cpp, gguf, pruning, reap, quantized, coding, tool-calling, text-generation, model-index, imatrix, conversational

Links

• Model: https://huggingface.co/crucible-labs/Qwen3.6-35B-A3B-REAP-48-v2-GGUF
• Files: https://huggingface.co/crucible-labs/Qwen3.6-35B-A3B-REAP-48-v2-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download crucible-labs/Qwen3.6-35B-A3B-REAP-48-v2-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.