OPEN LINK ↗
// 45d agoOPENSOURCE RELEASE
Qwen3-Coder 30B goes open, rivals Claude Sonnet
Alibaba's Qwen team open-sourced Qwen3-Coder, a family of agentic coding models including a 30B-A3B MoE variant (3B active parameters) under Apache 2.0 — delivering Claude Sonnet-level agentic coding performance at a fraction of the inference cost. The flagship 480B-A35B sets new open-model SOTA on SWE-Bench Verified, agentic browser-use, and tool-use benchmarks.
// ANALYSIS
An open-source MoE model that activates only 3B parameters yet competes with proprietary coding giants is a forcing function that reshapes the self-hosted AI stack overnight.
- –The 30B-A3B variant fits in consumer VRAM and runs locally — democratizing agentic coding without API costs or data privacy tradeoffs
- –Trained on 7.5T tokens (70% code) with execution-driven RL across 20,000 parallel environments: this is serious infrastructure-grade training discipline, not a finetune
- –Ships with Qwen Code, an open-source CLI agent (adapted from Claude Code) — Alibaba is competing at the tooling layer, not just the model layer
- –256K native context (extendable to 1M) makes it viable for repo-scale tasks where smaller context windows fall apart
- –Apache 2.0 license removes the commercial friction that limits Meta's Llama for enterprise deployment
// TAGS
qwen3-coderllmai-codingagentopen-weightsopen-sourceinferencemcp
DISCOVERED
45d ago
2026-04-16
PUBLISHED
45d ago
2026-04-16
RELEVANCE
9/ 10
AUTHOR
cmitsakis