AICodeKing

clem 🤗

Hacker News

<p>Alibaba has previewed Qwen 3.8 Max, a 2.4 trillion parameter model it’s calling second only to Anthropic’s Fable 5. The model is available now through Alibaba’s Token Plan, Qoder, QoderWork, and Qwen Chat. Open weights are confirmed but no release date or license terms have been given.</p>
<h2>The claim vs. the evidence</h2>
<p>The announcement landed a few days after rival Moonshot released Kimi K3 — complete with a benchmark table, pricing, and a hard open-weight date of July 27. Alibaba’s launch, by contrast, was a single post on X with no benchmark scores, no model card, and no architecture details.</p>
<p>AI analyst Julien Simon has pointed out the obvious problem: an undated, unverifiable ranking claim functions more as a narrative play than a model launch. There’s also a structural awkwardness here — Alibaba holds roughly a 36% stake in Moonshot, so publishing benchmarks that directly compare the two would be uncomfortable either way.</p>
<p>That said, Qwen’s two previous flagships shipped API-only, quietly breaking the lab’s open-weight tradition. Promising open weights for 3.8 Max is at least a signal they’re aware of the optics.</p>
<h2>Independent testing</h2>
<p>One detailed third-party evaluation tested Qwen 3.8 Max across eight tasks: 3D rendering, SVG generation, math reasoning, and agentic workflows. It scored 65/80 (81.25%), placing second overall — behind Fable 5 and ahead of Claude Opus 4. It got perfect scores on a bow-and-arrow game, a hard permutation problem, and a long-horizon fine-tuning task.</p>
<p>Separately, AI/ML API ran a Three.js benchmark where each model was given a single file containing a war galley, a Trojan horse, and a Greek war helmet, then asked to build three scenes in one shot. Qwen 3.8 Max finished considerably faster than the other models with no visible drop in output quality.</p>
<p>The one consistent weakness flagged in early testing was unreliable behavior during long agentic sessions in Claude Code. A subsequent update appears to have addressed this, along with improvements to front-end code generation and tool calling reliability.</p>
<h2>Pricing and access</h2>
<p>The model is currently available at 10% of standard pricing on Alibaba’s Token Plan. On Qoder, a 90% discount drops the billing coefficient to 0.05x, and off-peak hours (10pm–8am Singapore time) reduce it further to 0.01x — roughly 50x more usage for the same spend. A 14-day pro trial with 300 credits is also available.</p>
<p>Qoder also quietly launched a mystery model called “Cantus” priced at 3.2x the standard billing coefficient — about six times more expensive than Qwen 3.8 Max at standard rates — with no benchmarks or lab attribution. Speculation ranges from an unreleased Qwen Ultra variant to a frontier model from another lab being tested under a codename.</p>
<p>Qwen 3.8 Max is now free on Qwen Studio, though the free tier for Qwen Code CLI hosted inference has been discontinued.</p>
<h2>What’s still missing</h2>
<p>The open weights. Until those ship, the “second only to Fable 5” claim stays unverifiable. The model’s predecessor, Qwen3.7-Max, was the highest-ranked Chinese model on ECI, so a 2.4T successor being competitive at the frontier isn’t implausible — but implausible and unverified are different problems.</p>


Collections

Alibaba's Qwen team has announced Qwen3.8 Max, a 2.4 trillion parameter mixture-of-experts model positioned as frontier-competitive. A preview is live but open weights aren't released yet. Across 8 benchmark tasks including 3D rendering, SVG generation, math reasoning, and agentic workflows, it scored 65/80 (81.25%), placing second behind Fable 5 and ahead of Claude Opus 4. It achieved perfect scores on three tasks. A notable limitation is inconsistent behavior during long-running agentic sessions, particularly with Claude Code. The open weights release is confirmed but pending.

Qwen3.8 Max: Alibaba's 2.4 trillion parameter open-weight model benchmarked

<p>Even selling my two kidneys won’t be enough to run it locally, I guess!</p>


<p>Beating Claude Opus 4 on benchmarks as an open weight model is wild but the note on inconsistent performance during long claude code sessions shows once again that multi turn agentic stability is the real benchmark developers actually care about</p>


<p>Closed models will always be superior - just logically. There’s no reason to open source something that’s superior. The second a developed model becomes superior, they’ll close it. Such is capitalism.</p>


<p>What is seems to me with all those criticism over chinese models is that Americans keep to say it is not safe/legit/propaganda/propaganda/ etc etc to save a really big business for them. This is not safe, not an open weight model, baby!!!</p>
