Qwen3

Alibaba's open-weight large language model family launched April 2025 and evolved through multiple 2026 releases. Range covers six dense models (0.6B–32B parameters) and two Mixture-of-Experts models (30B-A3B and 235B-A22B at launch; Qwen 3.5 escalated to 397B-A17B with native vision + 201 languages in February 2026; Qwen3.6-27B added a hybrid Gated DeltaNet linear-attention + traditional-self-attention architecture in April 2026; Qwen3.7-Max preview ships 1M-token context). All releases under Apache 2.0 with downloadable weights.

Through 2025, the open-weight frontier was dominated by aggressively-sparse MoE designs (DeepSeek V3, Kimi K2) — and Alibaba's Qwen team was specifically targeting the same agentic-coding workload class. The 2026 evolution diverged: Qwen3.6-27B is dense (not MoE) but uses hybrid linear+traditional attention to match much-larger MoE models on agentic coding. The framing: density + hybrid attention can compete with sparse MoE at the workload class where most enterprise spend lives (agentic coding, retrieval-augmented chat).

Agentic coding (Qwen3-Coder-Next specifically optimizes for this workload), multilingual workloads (Qwen 3.5 supports 201 languages natively), long-horizon tasks via Qwen3.7-Max's 1M-token context + extended-thinking mode, on-prem deployment via the Apache-2.0 weights, and migration target for teams moving off proprietary frontier APIs.