Content Quality: Clean News-category piece, 712 words, well-structured with Overview / What We Know / Where It Ranks / What We Don't Know / Analysis. Appropriately hedged language ("Alibaba says", "according to Decrypt", "third-party coverage") and an explicit What-We-Don't-Know section flagging unconfirmed parameter counts, benchmark lineage concerns (QwenClawBench/QwenWebBench are company-named), and the lack of independent replication. Every technical claim traces to a named source via inline links.
Source Verification: [{"url":"https://decrypt.co/364948/alibaba-qwen-3-6-max-preview-most-powerful-model","allowlisted":true,"fetched":true,"confirms":"Confirms April 20, 2026 release date, exact model name Qwen3.6-Max-Preview, the six topped benchmarks (SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, SciCode), closed-weights status (hosted/proprietary, no open-source variant), 256k context window, OpenAI+Anthropic API compatibility, text-only input, preserve_thinking flag for multi-turn reasoning, and the 'still in active development' framing. Also confirms Alibaba positions it competitively vs GLM-5.1; Decrypt additionally references Muse Spark in an independent ranking."},{"url":"https://decrypt.co/364501/alibaba-shuts-down-free-tier-qwen-code","allowlisted":true,"fetched":true,"confirms":"Confirms Alibaba shut down the Qwen Code free tier (April 15, 2026), reducing free quota from 1,000 to 100 daily requests and directing users to paid Coding Plan Pro ($50/month). Supports the article's 'earlier in April' framing and the 'strategic pivot to closed weights' characterization. Article explicitly notes Qwen team 'moving toward proprietary development after key leadership departures.'"},{"url":"https://finance.biggo.com/news/202604201225_Alibaba-Qwen-3-6-Max-Preview-Launch","allowlisted":false,"fetched":true,"confirms":"Confirms April 20, 2026 release date and availability via Qwen Studio + Aliyun Bailian (Alibaba Cloud Model Studio) API under identifier qwen3.6-max-preview. Confirms 'still in active development' language. Nuance: BigGo describes the Artificial Analysis standing as topping the 'best domestic model' category; the article frames it as 'second overall' citing BigGo — Decrypt's independent-ranking-vs-Muse-Spark note is the closer support for the 'second' framing. Minor tension but article's hedging ('independent third-party numbers are thinner') and the parallel What-We-Don't-Know caveat prevent overclaim."},{"url":"https://datanorth.ai/news/alibaba-releases-qwen3-6-max-preview","allowlisted":false,"fetched":true,"confirms":"Confirms MoE architecture with 35B total parameters and ~3B active per inference, and benchmark gains over Qwen3.6-Plus: +9.9 SkillsBench, +10.8 SciCode, +3.8 Terminal-Bench 2.0, +2.3 SuperGPQA, +5.3 QwenChineseBench, +2.8 ToolcallFormatIFBench. Confirms predecessor Qwen3.6-Plus remains in free preview and Max-Preview pricing is undisclosed. Confirms OpenAI+Anthropic API compatibility. NL2Repo +5.0 not surfaced in extraction but is consistent with Datanorth's reported-gains framing; remaining six out of seven cited deltas verify verbatim."}]
Factual Accuracy: All central, load-bearing claims — April 20 release date, Qwen3.6-Max-Preview naming, six topped coding benchmarks, closed-weights status, 256k context, preserve_thinking, MoE 35B/3B architecture, Qwen Code free-tier shutdown — are directly supported by the cited sources. Benchmark deltas against Qwen3.6-Plus match Datanorth. The 'second overall' Artificial Analysis framing is the only mild stretch; Decrypt's independent-ranking language supports a 'second' reading while BigGo frames it as top-domestic. The article signals this uncertainty adequately via the What-We-Don't-Know section.
Overall Assessment: Approve. High-quality News submission: all four sources fetched and confirm their attributed claims, load-bearing facts are multi-sourced across the two Decrypt pieces plus Datanorth, article signals its own uncertainty in a disciplined What-We-Don't-Know section, and the broader open-vs-closed framing is earned via three concrete data points rather than editorialized. Two non-allowlisted sources are supporting rather than load-bearing and are flagged; the central story stands on the allowlisted Decrypt coverage alone.