Qwen3-Coder-Next
Frontier-adjacent coding from 3B active parameters
What it is
Alibaba's coding MoE: 80B total parameters with only 3B active, a hybrid Gated DeltaNet + gated attention layout across 512 experts, and 262K context natively. Scores 70.6% on SWE-Bench Verified and 44.3% on SWE-Bench Pro — territory that recently demanded 10–20x the active compute.
Why it's interesting
The economics of agentic coding, moved. Apache-2.0 weights, a million Hugging Face downloads a month, and an activation footprint small enough that quantized builds run serious agent workloads on a single big GPU.
Use cases
- Self-hosted agentic coding behind vLLM or SGLang
- Long-context repository-scale refactoring
- Fine-tuning a coding specialist you actually own
Who it's for
AI engineers, self-hosting developers, teams leaving API coding models
Setup
Moderate. 159GB of weights at BF16 — multi-GPU or a heavily quantized GGUF for single-card use. Served via vLLM, SGLang, or llama.cpp
Limitations & cautions
Full-precision inference is multi-GPU territory, the hardest agentic benchmarks still favor closed flagships, and the missing LICENSE file is a paperwork wobble worth watching.
Editorial takeaway
Every few months the compute floor for real coding agents drops through another story. This is the current basement.