Qwen3-Coder-Next MoE GGUF
Certified GGUF precision quantization ladder for Qwen/Qwen3-Coder-Next hybrid Gated DeltaNet and mixture-of-experts coding model.
Measured GGUF precision variants (Q8_0, Q4_K_M, Q2_K) of Qwen/Qwen3-Coder-Next (79.6B total parameters, 3.14B active per token across 512 routed + 1 shared experts with top-10 routing and 262K context) evaluated across standard reasoning, mathematics, and science benchmarks (ARC-MC, GSM8K, MATH-500) under reproducible seed controls.
Architecture
- Parameters
- 79.60B
- Active
- 3.14B
- Layers
- 48
- Quantization ladder
- Certified GGUF
- Experts
- 512
- Experts before
- —
- Top-k
- 10
Downloads
Exact release files
3 immutable artifacts
Qwen3-Coder-Next.q4_k_m.gguf
Size: 48 GB
64 GB+ unified memory / multi-GPU or CPU+GPU MoE offload (~70% compression vs BF16). Retains 100% GSM8K accuracy (95.5%), 97.0% ARC-MC (+0.5 pp), and 61.0% MATH-500 (-0.5 pp).
SHA-256 Checksum
c6a4c85791cf23606278e3131a0d723d3f720c75d51f4624d99a33e7cff1d6db
Qwen3-Coder-Next.q8_0.gguf
Size: 85 GB
96–128 GB+ unified memory or multi-GPU workstation. Near-lossless precision across ARC-MC (96.5%), GSM8K (95.5%), and MATH-500 (62.5%, +1.0 pp vs BF16).
SHA-256 Checksum
05dc5947082de18697657abb42f2613ebe9c149e0cf88a8d0a375f786a4205db
Qwen3-Coder-Next.q2_k.gguf
Size: 29 GB
32–48 GB reference capacity (~82% compression). Preserves 96.5% ARC-MC and 91.5% GSM8K (-4.0 pp); note MATH-500 degradation (-13.5 pp).
SHA-256 Checksum
df7a317496f362369573e3ac7cdc22742f4d24f0873f4ad89942bd9eac3510f1
Deployment
Quickstart & CLI
llama.cpp & Ollama
Stream weights directly from Hugging Face via Ollama CLI:
ollama run hf.co/Omega-Factory/Qwen3-Coder-Next-GGUF:Q4_K_MRun inference locally using llama.cpp with the recommended artifact:
llama-cli -hf Omega-Factory/Qwen3-Coder-Next-GGUF -m Qwen3-Coder-Next.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-Coder-Next-GGUF -m Qwen3-Coder-Next.q4_k_m.gguf --port 8080 -ngl 99Evidence
Measured outcomes & precision ladder
Precision & Accuracy Matrix vs. BF16 Baseline
Holdout evaluation split (N=200 per category, seed 42) comparing all quant variants to unquantized BF16 baseline.
ARC-MC
96.5%
GSM8K
95.5%
MATH-500
61.5%
ARC-MC
96.5%
GSM8K
95.5%
MATH-500
62.5%
+1.0 pp
ARC-MC
97.0%
+0.5 pp
GSM8K
95.5%
MATH-500
61.0%
-0.5 pp
ARC-MC
96.5%
GSM8K
91.5%
-4.0 pp
MATH-500
48.0%
-13.5 pp
Item-Level McNemar Discordance & Flips vs. BF16 Baseline
Per-item analysis measuring individual test questions that flipped answer status relative to the unquantized BF16 baseline (N=200).
Recommended Release Rung Scoped Verification (Q4_K_M)
Formal cryptographic release marker evaluation records with 95% confidence intervals and sample counts.
ARC-Challenge Science Reasoning
qwen3-coder-next-arc-mc-v1 · accuracy_pct
Base
96.50%
Release
97.00%
Delta
+0.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under standard multiple-choice log-likelihood scoring.
95% CI [0.00, 1.50] pp · 200 samples · 1 seeds · artifact c6a4c85791cf23606278e3131a0d723d3f720c75d51f4624d99a33e7cff1d6db
GSM8K Grade School Math
qwen3-coder-next-gsm8k-v1 · accuracy_pct
Base
95.50%
Release
95.50%
Delta
+0.00 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under greedy decoding.
95% CI [-2.00, 2.00] pp · 200 samples · 1 seeds · artifact c6a4c85791cf23606278e3131a0d723d3f720c75d51f4624d99a33e7cff1d6db
MATH-500 Advanced Mathematics
qwen3-coder-next-math500-v1 · accuracy_pct
Base
61.50%
Release
61.00%
Delta
-0.50 pp
Holdout evaluation set (N=200, seed 42) comparing Q4_K_M quantized release against uncompressed BF16 baseline under greedy decoding.
95% CI [-5.50, 5.00] pp · 200 samples · 1 seeds · artifact c6a4c85791cf23606278e3131a0d723d3f720c75d51f4624d99a33e7cff1d6db
Literature & Architecture
Foundational Research & Citations
Primary scientific publications & architecture preprints
Qwen3 Technical Report
Qwen Team, Alibaba Group
Introduces the Qwen3 foundation model suite and fine-grained Mixture-of-Experts architecture powering Qwen3-Coder-Next (80B total, 3B active parameters across 512 routed + 1 shared experts with 256K context for agentic coding).
Gated Delta Networks: Improving Mamba2 with Delta Rule
Songlin Yang, Jan Kautz, Ali Hatamizadeh
Introduces Gated DeltaNet linear attention combining gating and delta-update memory rules, forming the 3:1 hybrid Gated DeltaNet + Gated Attention backbone of Qwen3-Coder-Next.
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
- HYBRID_DELTANET_MOE_ARCHITECTURE — Hybrid architecture combining 36 Gated DeltaNet linear attention blocks and 12 gated full-attention blocks with 512 routed experts (top-10 active) + 1 shared expert and 262,144-token native context window.
- VRAM_ESTIMATE_NOTICE — Artifact size is disk usage, not a measured VRAM requirement. Deployment requires peak-memory profiling.
- COMPRESSED_DERIVATIVE — Quantized GGUF precision ladder of Qwen/Qwen3-Coder-Next under Apache 2.0 license.
Public release identity
- Release ID
- qwen3-coder-next-gguf-v1
- Release version
- 1
- Published
- 2026-10-14T07:53:45.000Z
- Model ID
- Omega-Factory/Qwen3-Coder-Next-GGUF
- Base model ID
- Qwen/Qwen3-Coder-Next
- Catalog family
- Not supplied
- Quantization rung
- Not supplied
- Supersedes
- Not supplied
- HF repository
- Omega-Factory/Qwen3-Coder-Next-GGUF
- HF repository URL
- https://huggingface.co/Omega-Factory/Qwen3-Coder-Next-GGUF
- HF revision
- 9a9112af264ac5baef1b73dea5fc89b5ca632fb3
- HF revision URL
- https://huggingface.co/Omega-Factory/Qwen3-Coder-Next-GGUF/tree/9a9112af264ac5baef1b73dea5fc89b5ca632fb3
- Release manifest SHA-256
- 1833a4323a3638d276869529124c8c122c0d5a755dfa22a6af422dcd39e670b7
- Catalog SHA-256
- d853e89023b5b0378c96b28cf519639e2b0bb9b774d171f1fe1e0dbcc48b68f8
- Catalog generated
- 2026-10-14T08:00:00Z
- License
- Apache 2.0
- License ID
- apache-2.0
- Commercial use
- ALLOWED
- License URL
- https://www.apache.org/licenses/LICENSE-2.0