Qwen promised "open source to the end" 12 days ago, then shipped closed-weight Qwen4-Max
Delicious_Character6 · reddit · 2026-11-18
A Reddit post points out that Alibaba's Qwen team reaffirmed "open source to the end" just 12 days ago, then launched its flagship MoE Qwen4-Max as a closed-weight API:
- Architecture: 2.4T total params, 71B active per token, 1M context; Gated DeltaNet (GDN) layers plus Qwen Sparse Attention (QSA) cut active params from 95B to 71B and task-level inference cost by 60%.
- The catch: zero downloadable weights — no HF repo, no safetensors, no GGUF quants — locked entirely behind QwenCloud API at $1.50 input / $4.50 output per million tokens.
- Author's take: a familiar playbook — build community goodwill, harvest free benchmarks and fine-tuning tooling, then slam the vault shut at leaderboard parity. For self-hosters, if it can't run in vLLM or llama.cpp, it's just another closed endpoint with no fallback.
(Unverified forum claim; treat specifics with caution.)
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22