Qwen3.5 9B, self-quantized to GGUF by Atomic Chat. Built straight from Qwen's original weights with a per-tensor importance matrix, so this is not a repack of somebody else's files. Runs fully offline.
Highlights
- 9.7B parameters: the weights this repo quantizes.
- Context length: 262,144 tokens (256K), as published by Qwen.
- 32 layers: Dense decoder.
- Modalities: the base model handles Text, Image; this repo ships text-only quants, it carries no vision projector.
- Full imatrix ladder: every quant is calibrated with an importance matrix.
- Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.
- Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
[!NOTE] These GGUFs are self-quantized from the original weights, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.
[!IMPORTANT] Always pass
--jinjaso the Qwen3.5 9B chat template is applied. Without it the model can emit malformed turns.
Model Overview
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3.5-9B |
| Parameters | 9.7B |
| Layers | 32 |
| Context length | 262,144 tokens (256K) |
| Vocabulary | 248,320 |
| Modalities | Text, Image in the base model; text only in this repo, it ships no vision projector |
| Architecture | Dense decoder, 16 attention heads over 4 KV heads, Qwen3_5ForConditionalGeneration |
| This repo | GGUF quants (imatrix). Quants: Q4_K_M, UD-Q4_K_XL, Q5_K_M, Q6_K, Q8_0 |
Scores are Qwen's published results for the base Qwen/Qwen3.5-9B, not our own measurements. Quantization preserves the large majority of this; Q4_K_M and up stay close to full precision.
Choosing a quant
| Quant | Size | Notes |
|---|---|---|
Q4_K_M | 5.6 GB | Recommended default. Best balance of size, speed and quality. |
UD-Q4_K_XL | 6.4 GB | Dynamic. Embeddings and output kept at Q8_0 for higher quality at a Q4 footprint. |
Q5_K_M | 6.5 GB | Higher quality, low loss. |
Q6_K | 7.4 GB | Near lossless, noticeably lighter than Q8_0. |
Q8_0 | 9.5 GB | Effectively lossless, reference quality. |
[!TIP] Pick the largest file that fits your (V)RAM with room for context.
Q4_K_MorUD-Q4_K_XLis the sweet spot for most setups;Q6_KorQ8_0for maximum fidelity.
Get started
Run Qwen3.5 9B locally with:
- Atomic Chat: the easiest path. Open the app, search
AtomicChat/Qwen3.5-9B-GGUF, pick a quant, hit Use this model. - llama.cpp:
llama-server -hf AtomicChat/Qwen3.5-9B-GGUF:Q4_K_M --jinja -c 8192 - Ollama:
ollama run hf.co/AtomicChat/Qwen3.5-9B-GGUF:Q4_K_M - LM Studio / Jan: search the repo id, download any quant.
Best practices
| Parameter | Value |
|---|---|
| temperature | 1.0 |
| top_p | 0.95 |
| top_k | 20 |
| min_p | 0.0 |
| repetition_penalty | 1.0 |
Qwen's recommended sampling configuration for Qwen/Qwen3.5-9B.
Run in llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server
./llama.cpp/build/bin/llama-server \
-hf AtomicChat/Qwen3.5-9B-GGUF:Q4_K_M \
--jinja -ngl 99 -c 8192 -fa on
How these were made
- Download
Qwen/Qwen3.5-9B(original weights). - Convert to f16 GGUF with llama.cpp.
- Build an importance matrix over our calibration corpus.
- Quantize the ladder with
--imatrix. UD-Q4_K_XLadditionally pins the token-embedding and output tensors toQ8_0.
License
Original model by Qwen, released under the Apache 2.0 license. Full terms: Apache 2.0. Quantized by Atomic Chat.