Qwen3-Coder-Next
Qwen3-Coder-Next is a programming-specialized foundation model developed by Alibaba. With 80B parameters and a long 256K context window, it enables advanced code generation and analysis.
파라미터
80.0B
컨텍스트
256K
라이선스
https://huggingface.co/Qwen/Qwen2.5-72B/blob/main/LICENSE
출시일
2026-02-03
API 가격
이 모델의 API 가격 정보는 현재 공개되지 않았습니다
강점
- ・80B strong parameter scale
- ・256K extensive context window
- ・Advanced programming capabilities
약점
- ・Closed license model
- ・Large model size up to 48GB
- ・High computational resource requirement
활용 사례
- ・Large codebase analysis
- ・Auto-generating complex features
- ・Advanced programming assistance
심층 분석
Release Date
February 2026
Total Parameters
80B
MoE with 512 experts
Active Parameters
3B per token
10 experts + 1 shared active
Layers
48
Hidden Dimension
2,048
Context Window
262,144 tokens
Architecture
Hybrid: Gated DeltaNet + Gated Attention + MoE
Input Price
$0.07-$0.12/1M tokens
Output Price
$0.30-$0.75/1M tokens
SWE-bench Verified
70.6
Competitive with DeepSeek-V3.2 (70.2) at 671B
강점
- ・SWE-bench 70.6 at only 3B active parameters — competitive with 671B DeepSeek-V3.2 (70.2)
- ・Incredibly cheap: $0.07/$0.30 on Alibaba Cloud, $0.12/$0.75 on third-party providers
- ・Trained on 800K verifiable executable tasks with RL for long-horizon agent coding
- ・256K context for keeping entire codebases in a single session
- ・Apache 2.0 licensed — available for self-hosting and fine-tuning
약점
- ・Non-thinking mode only — does not support blocks
- ・SWE-bench Pro 44.3% leaves significant room for improvement on harder tasks
- ・Terminal-Bench 36.2% is notably lower than larger coding models
- ・80B total parameters still require ~46GB VRAM at 4-bit quantization
- ・Specialized for coding — may underperform on general knowledge tasks
경쟁사 비교
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| DeepSeek-V3.2 | ~1430 | 70.2 | ~82 | $0.14/$0.28 |
| GLM-4.7 | ~1420 | 74.2 | ~80 | Proprietary |
| Qwen3-Coder-Next | ~1400 | 70.6 | ~75 | $0.07/$0.30 |
| Claude Sonnet 4.6 | ~1470 | ~78 | ~87 | $3.00/$15.00 |
| GPT-4o | ~1450 | ~72 | ~85 | $2.50/$10.00 |
Qwen3-Coder-Next is an open-weight coding-specialized model from February 2026, built on the Qwen3-Next-80B-A3B backbone with 80B total parameters and only 3B active per token. It achieves SWE-bench Verified 70.6 — competitive with the 671B DeepSeek-V3.2 — at a fraction of the compute cost. Trained on 800K executable tasks with reinforcement learning, it is purpose-built for agentic coding workflows, IDE copilots, and long-horizon development tasks.
출처
분석 생성일: 2026-05-24