Parameters
0.1K total / 0.0K active
MoE: total / active
Architecture
Causal Language Model
Released
30.01.2026
License
Apache License 2.0
Input Modalities
Output Modalities
Context (native)
262,144 tokens
Context (extended)
262,144 tokens
About
Qwen3-Coder-Next (Qwen/Qwen3-Coder-Next) is Alibaba's open-weight language model designed specifically for coding agents and local development - a sparse Mixture-of-Experts with 80B total parameters and only 3B activated (79B non-embedding), delivering performance comparable to models with 10-20x more active parameters at high cost-effectiveness for agent deployment.
Its 48-layer hybrid layout repeats 12x (3x (Gated DeltaNet -> MoE) + 1x (Gated Attention -> MoE)): 36 linear-attention layers (Gated DeltaNet with 32 V / 16 QK heads, head dim 128) and 12 full-attention layers (Gated Attention with 16 Q / 2 KV heads, head dim 256, RoPE dim 64), with hidden size 2048. The MoE uses 512 experts with 10 activated per token plus 1 shared expert (expert intermediate dim 512). Context length is 262,144 tokens natively.
Qwen3-Coder-Next excels at long-horizon reasoning, complex tool usage and recovery from execution failures via an elaborate agentic training recipe, and integrates seamlessly with real-world IDE/CLI scaffolds (Claude Code, Qwen Code, Qoder, Kilo, Trae, Cline and more). Note: it supports non-thinking mode only - it does not generate <think> blocks, and enable_thinking=False is no longer required. Released January 30, 2026 under Apache 2.0.
Training Data Pretraining & Post-training
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
SWE-bench Verified
coding_agent
|
80.50%
|
03.02.2026 |
#38/60
better than 37%
|
|
SWE-bench Pro
coding_agent
|
49.39%
|
03.02.2026 |
#55/77
better than 29%
|
|
Terminal-Bench 2.0
coding_agent
|
11.75%
|
03.02.2026 |
#20/23
better than 14%
|
|
Aider
coding_agent
|
79.21%
|
03.02.2026 |
#2/4
better than 67%
|
|
SWE-bench Multilingual
coding_agent
|
68.40%
|
03.02.2026 |
#35/48
better than 28%
|
What is a good SWE-bench Verified score? What is a good SWE-bench Pro score? What is a good Terminal-Bench 2.0 score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 0.0K are active per token
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 0.0 GB | 0.0 GB | 0.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 0.0 GB | 0.0 GB | 0.0 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 0.0 GB | 0.0 GB | 0.0 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 0.0 GB | 0.0 GB | 0.0 GB | Which GPUs can run it? Calculate on your GPU |
Consumer GPUs that fit this model with a single card (INT4):
Architecture
- Attention
- Hybrid Attention (16:2)
- MoE
- 512 experts · top-10 per token
- Layers
- 48
- Hidden size
- 2048
- Context
- 262K tokens
- RoPE θ
- 5M
- Parameters
- 80
- Active params
- 3
Source: Hugging Face config.json · Qwen3NextForCausalLM · exact layer pattern · model repo
Training Pipeline
-
1
pretraining
Pretraining
Pretraining stage with large-scale text and code corpora
-
2
other
Post-training
Post-training stage including elaborate training recipe for agentic capabilities, long-horizon reasoning, complex tool usage, and recovery from execution failures
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| Coding-agent training corpus (long-horizon agentic coding) | finetune | — | — |
Linked Resources
Qwen3-Coder-Next Blog
https://qwen.ai/blog?id=qwen3-coder-next
Qwen3-Coder GitHub Repository
https://github.com/QwenLM/Qwen3-Coder
Qwen3-Coder-Next Technical Report
https://github.com/QwenLM/Qwen3-Coder/blob/main/qwen3_coder_next_tech_report.pdf
Qwen3 Documentation
https://qwen.readthedocs.io/en/latest/
Trend Analysis
24h Change
+0.2%
7d Change
+1.6%
30d Change
+5.6%
Current
116,019
downloads
+0.0%
likes
+0.1%
downloads_all_time
+0.1%
downloads
-0.3%
No comments yet
Be the first to share your thoughts!