Qwen 3.x → 3.5/3.8

As of: 2026-09 (3.8-Omni-Flash era)

The line

• Qwen3 (Apr 2025): 0.6B to 235B-A22B MoE; hybrid thinking — one model switches fast/thinking modes (/think, /no_think); Apache-2.0; 119 languages.
• Qwen3.5 → 3.8 (2026): incremental capability + Omni variants (native text/audio/vision in open weights) and speed-focused "Flash" tiers.
• Consistently the top of open leaderboards per size class through the period; the most downloaded model family on HF in 2026 stretches.

Why it matters operationally

• Apache-2.0 — genuinely OSI, usable anywhere including platforms (unlike Llama's custom license).
• Every size tier exists (0.6B→MoE); one family covers edge (phone/laptop) to server.
• Fine-tuning ecosystem (unsloth recipes etc.) is first-class; community quants (MLX/GGUF) drop within days.

Default recommendation for local/self-hosted agents in 2026: benchmark Qwen current vs Gemma vs Mistral at your task; Qwen wins more often than not on code+multilingual.