Qwen3 Coder 30B A3B, self-quantized to GGUF by Atomic Chat. Built straight from Qwen's original weights with a per-tensor importance matrix. Runs fully offline.
Highlights
- Agentic coding specialist with significant performance among open models on agentic coding, agentic browser-use, and other foundational coding tasks.
- Efficient MoE: 30.5B total parameters, only 3.3B activated per token (128 experts, 8 activated).
- 256K native context (262,144 tokens), extendable up to ~1M tokens with Yarn, optimized for repository-scale understanding.
- Tool calling built in with a specially designed function-call format, supporting platforms such as Qwen Code and CLINE.
- Non-thinking mode only — does not emit
<think></think>blocks; noenable_thinkingflag required. - Full quant ladder with an importance matrix on every quant over
calibration_datav3.
[!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 Coder 30B A3B chat template is applied. Without it the model can emit malformed turns.
Model Overview
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3-Coder-30B-A3B-Instruct |
| Total / active parameters | 30.5B total, 3.3B activated (128 experts, 8 activated) |
| Layers | 48 |
| Context length | 262,144 native (extendable to ~1M with Yarn) |
| Architecture | Causal LM, Mixture-of-Experts; GQA (32 Q heads, 4 KV heads) |
| This repo | GGUF quants (imatrix) |
See the official model card for Qwen's published benchmark results.
Choosing a quant
| Quant | Size | Notes |
|---|---|---|
Q2_K | 11.3 GB | Smallest. Minimal RAM, clear quality drop. |
IQ3_M | 13.5 GB | Beats Q3 at similar size thanks to imatrix. Best low-RAM pick. |
Q3_K_M | 14.7 GB | Low quality but usable. |
Q3_K_L | 15.9 GB | A step above Q3_K_M. |
IQ4_XS | 16.4 GB | Excellent quality for size. Recommended low-bit. |
Q4_K_S | 17.5 GB | Compact Q4, fast. |
Q4_K_M | 18.6 GB | Recommended default. Best balance of size, speed and quality. |
UD-Q4_K_XL | 18.8 GB | Dynamic. Embeddings and output kept at Q8_0 for higher quality at a Q4 footprint. |
Q5_K_S | 19.7 GB | Higher quality. |
Q5_K_M | 12.1 GB | Higher quality, low loss. |
Q6_K | 17.4 GB | Near lossless. |
Q8_0 | 20.3 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 Coder 30B A3B locally with:
- Atomic Chat: the easiest path. Open the app, search
AlexAtomic/qwen3-coder-30b-a3b-GGUF, pick a quant, hit Use this model. - llama.cpp:
llama-server -hf AlexAtomic/qwen3-coder-30b-a3b-GGUF:Q4_K_M --jinja -c 8192 - Ollama:
ollama run hf.co/AlexAtomic/qwen3-coder-30b-a3b-GGUF:Q4_K_M - LM Studio / Jan: search the repo id, download any quant.
Best practices
| Parameter | Value |
|---|---|
| temperature | 0.7 |
| top_p | 0.8 |
| top_k | 20 |
| repetition_penalty | 1.05 |
Qwen's recommended settings for this model (non-thinking); recommended output length 65,536 tokens.
Run in llama.cpp
git clone https://github.com/ggerganov/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 AlexAtomic/qwen3-coder-30b-a3b-GGUF:UD-Q4_K_XL \
--jinja -ngl 99 -c 8192 -fa on
How these were made
- Download
Qwen/Qwen3-Coder-30B-A3B-Instruct(original weights). - Convert to f16 GGUF with llama.cpp.
- Build an importance matrix over
calibration_datav3(100 chunks). - Quantize the full ladder with
--imatrix. UD-Q4_K_XLadditionally pins the token-embedding and output tensors toQ8_0.
License
Released by Qwen under the Apache 2.0 license. Quantized by Atomic Chat.