Qwen3-30B-A3B GGUF
Certified GGUF precision quantization ladder for Qwen/Qwen3-30B-A3B pruned and healed MoE architecture.
Certified GGUF precision quantization ladder for Qwen/Qwen3-30B-A3B pruned and healed MoE architecture (96 local experts, 8 active per token) across Q8_0, Q4_K_M, and Q2_K.
Architecture
- Parameters
- 23.28B
- Parameters before
- 30.50B
- Active
- 3.30B
- Layers
- 48
- Compression
- 23.7%
- Experts
- 96
- Experts before
- 128
- Top-k
- 8
Downloads
Exact release files
3 immutable artifacts
Qwen3-30B-A3B.q4_k_m.gguf
Size: 14 GB
20–24 GB VRAM (RTX 3090 / 4090 / Mac M 32GB). General conversational and daily development.
SHA-256 Checksum
ebd160a8a0e47cf3ea0a6b4738b07f2017a78acc829f6147977009e7a9bc37aa
Qwen3-30B-A3B.q8_0.gguf
Size: 25 GB
32–48 GB VRAM. Production, critical reasoning, and coding.
SHA-256 Checksum
6c05ae5df8133f3cbb40a5e781a186d92dc487821a990ba47b7d268f3e3c4c5e
Qwen3-30B-A3B.q2_k.gguf
Size: 8.6 GB
12–16 GB VRAM (RTX 4070 / 4080 / Mac M 16GB). Memory-constrained consumer workstations.
SHA-256 Checksum
2776de7901063fd2639a79073c68028814aaf9393525bab57af407e6a53884c4
Deployment
Quickstart & CLI
llama.cpp & Ollama
Stream weights directly from Hugging Face via Ollama CLI:
ollama run hf.co/Omega-Factory/Qwen3-30B-A3B-GGUF:Q4_K_MRun inference locally using llama.cpp with the recommended artifact:
llama-cli -hf Omega-Factory/Qwen3-30B-A3B-GGUF -m Qwen3-30B-A3B.q4_k_m.gguf -p "Explain the core mechanics of mixture-of-experts models." -n 512Launch a local OpenAI-compatible HTTP server:
llama-server -hf Omega-Factory/Qwen3-30B-A3B-GGUF -m Qwen3-30B-A3B.q4_k_m.gguf --port 8080 -ngl 99Evidence
Measured outcomes & precision ladder
Precision & Accuracy Matrix vs. FP16 Baseline
Holdout evaluation split (N=200 per category, seed 42) comparing all quant variants to unquantized baseline.
ARC-MC
93.5%
GSM8K
92.5%
MATH-500
49.5%
ARC-MC
93.5%
GSM8K
94.0%
+1.5 pp
MATH-500
48.5%
-1.0 pp
ARC-MC
93.0%
-0.5 pp
GSM8K
94.0%
+1.5 pp
MATH-500
48.0%
-1.5 pp
ARC-MC
86.5%
-7.0 pp
GSM8K
87.5%
-5.0 pp
MATH-500
38.5%
-11.0 pp
Item-Level McNemar Discordance & Flips vs. FP16 Baseline
Per-item analysis measuring individual test questions that flipped answer status relative to the unquantized FP16 baseline (N=200).
Recommended Release Rung Scoped Verification (Q4_K_M)
Formal cryptographic release marker evaluation records with 95% confidence intervals and sample counts.
GSM8K Grade School Math
qwen3-30b-a3b-gsm8k-v1 · accuracy_pct
Base
92.50%
Release
94.00%
Delta
+1.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under greedy decoding.
95% CI [-3.42, 6.42] pp · 200 samples · 1 seeds · artifact ebd160a8a0e47cf3ea0a6b4738b07f2017a78acc829f6147977009e7a9bc37aa
MATH-500 Advanced Mathematics
qwen3-30b-a3b-math500-v1 · accuracy_pct
Base
49.50%
Release
48.00%
Delta
-1.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under greedy decoding.
95% CI [-11.30, 8.30] pp · 200 samples · 1 seeds · artifact ebd160a8a0e47cf3ea0a6b4738b07f2017a78acc829f6147977009e7a9bc37aa
ARC-Challenge Science Reasoning
qwen3-30b-a3b-arc-mc-v1 · accuracy_pct
Base
93.50%
Release
93.00%
Delta
-0.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed FP16 baseline under standard multiple-choice log-likelihood scoring.
95% CI [-5.42, 4.42] pp · 200 samples · 1 seeds · artifact ebd160a8a0e47cf3ea0a6b4738b07f2017a78acc829f6147977009e7a9bc37aa
Literature & Architecture
Foundational Research & Citations
Primary scientific publications & architecture preprints
Qwen3 Technical Report
Qwen Team, Alibaba Group
Introduces the Qwen3 foundation model suite across dense and Mixture-of-Experts (MoE) architectures, featuring unified thinking/non-thinking modes, dynamic reasoning budgets, and high-efficiency token routing.
Which Quantization Should I Use? A Unified Evaluation of llama.cpp Quantization
Empirical GGUF Research Study
Comprehensive empirical analysis of llama.cpp GGUF quantization schemes (Q8_0, Q4_K_M, Q2_K) across reasoning benchmarks, memory footprints, and CPU/GPU throughput.
Limitations
- PRUNED_MOE_ARCHITECTURE — 96 local experts active in MoE architecture with top-8 routing per token.
- VRAM_ESTIMATE_NOTICE — Artifact size is disk usage, not a measured VRAM requirement. Deployment requires peak-memory profiling.
- COMPRESSED_DERIVATIVE — Pruned and healed derivative of Qwen/Qwen3-30B-A3B under Apache 2.0 license.
Public release identity
- Release ID
- qwen3-30b-a3b-gguf-v1
- Release version
- 1
- Published
- 2026-10-10T13:00:59.000Z
- Model ID
- Omega-Factory/Qwen3-30B-A3B-GGUF
- Base model ID
- Qwen/Qwen3-30B-A3B
- Catalog family
- Not supplied
- Quantization rung
- Not supplied
- Supersedes
- Not supplied
- HF repository
- Omega-Factory/Qwen3-30B-A3B-GGUF
- HF repository URL
- https://huggingface.co/Omega-Factory/Qwen3-30B-A3B-GGUF
- HF revision
- f8ddac64b7dc5caed9091d40db8640f5eefe54fb
- HF revision URL
- https://huggingface.co/Omega-Factory/Qwen3-30B-A3B-GGUF/tree/f8ddac64b7dc5caed9091d40db8640f5eefe54fb
- Release manifest SHA-256
- 4e1bbef6c9053896dfa209b55589c36203cf0df07fcf5b80a133d1c448bb3f3e
- Catalog SHA-256
- e87595a7c9a5d0a6ae3662985e99356632737f2a673b773c938e4f93b214a99a
- Catalog generated
- 2026-10-11T12:45:00Z
- Commercial use
- ALLOWED
- License URL
- https://www.apache.org/licenses/LICENSE-2.0