Qwen3 Coder, about 71% cheaper.

Qwen3-Coder-Next is Alibaba’s open-weight coding-agent model, released on 2026-02-02 under Apache 2.0. It activates just 3 billion of its 80 billion parameters per token, yet lands SWE-bench numbers that took 10–20× more active compute a year earlier — which is why it became the default budget engine inside coding agents like Cline within days of launch.

The Qwen machine

Qwen is Alibaba Cloud’s model family, open-sourced since September 2023 and shipped at every size point under permissive licenses. The strategy produced the most-derived open family in the world: over 200,000 Qwen variants on Hugging Face by March 2026 (Wikipedia). Behind it sits real money — a RMB 380 billion (~$53B) three-year AI infrastructure plan that Alibaba’s CEO has since said the company will exceed (CNBC).

A footnote with weight: within a month of this model shipping, the Qwen team lost three senior leaders — including coding lead Binyuan Hui, whose group built the Coder line, and overall tech lead Junyang Lin, who signed off with “bye my beloved qwen.” Alibaba responded with a restructuring and an internal AGI task force (CNBC). Coder-Next is effectively the last coder flagship of that era.

From 480B to 80B

The line runs: Qwen2.5-Coder (November 2024) → Qwen3 (April 2025) → Qwen3-Coder-480B-A35B (July 2025, the first big agentic coder, which Cerebras later served at up to 2,000 tokens/second) → Qwen3-Next-80B (September 2025, a new hybrid architecture) → Qwen3-Coder-Next (February 2026), which puts the coder recipe on the efficient Next architecture. It supports 358 programming languages and a 262,144-token native context (extendable to 1M with YaRN).

Why 3B active parameters is enough

• Hybrid attention: 48 layers alternating 36 linear-attention (Gated DeltaNet) layers with 12 gated full-attention layers — the trick that keeps quarter-million-token agent loops cheap.
• Ultra-sparse MoE: 512 experts, 10 + 1 shared active per token → 80B total, ~3B active.
• Non-thinking by design: it never emits think-blocks, which Qwen pitches as simpler and faster for production agents.
• Trained like an agent: tasks synthesized from real GitHub PRs, run in executable environments across six scaffolds (SWE-agent, OpenHands, Claude Code, Qwen Code and more), with a 480B teacher distilling domain experts back into one model.
• Local-first: ~46GB at 4-bit — the first agent-grade coder that fits a 64GB MacBook Pro, a point Simon Willison made in the 735-point Hacker News launch thread.

What the benchmarks say

Pricing: tiers vs flat

Alibaba’s own Model Studio bills Coder-Next in tiers by request size: $0.30/$1.50 per million tokens up to 32K input, $0.50/$2.50 to 128K, $0.80/$4.00 to 256K — and one big-context request bills *all* its tokens at the tier it lands in. Third-party hosting undercuts that (the cheapest hosts sit around $0.11/$0.80 after ninety days of price war), and adoption ran ahead of it: 1.58 billion prompt tokens through OpenRouter on day one, with coding agents pi, Hermes Agent and Cline its biggest public consumers. Smouter’s live rate right now: $0.081 in / $0.46 out flat — 71% under Alibaba’s lowest tier, whatever the request size.