ModelBrief · Qwen4-Max · Topic Page
Last updated Nov 20, 2026

OrcaRouter Day Two: 147s First Token, Vals 63.2%

Published Nov 17, 2026 · Last updated Nov 17, 2026

On day two, two gaps closed: the latency number arrived, and the leak story collapsed. OrcaRouter updated the same long piece on November 17 (originally a check of the vLLM leak) to test, in one place, the first measured latency, a second independent index score, the GA product page's spec corrections, and a same-day claim that a distillation crackdown is holding Chinese models back. The picture splits: the model is the top Chinese entry on a white-collar index, 16 points behind on a terminal-work index, and its ugliest number is how long the first token takes.

Latency lands: the first token is the bad part

On launch day AA had no time-per-task figure for this model. It has one now: 32.1 minutes per Intelligence Index task, against a 13.3-minute median for its price tier — "notably slow", in AA's words. Total task time is not the worst in the set, though: Kimi K3.1 Pro takes 38.8 minutes and Claude Fable 5.2 takes 40.8, both longer; Opus 5.5 takes 24.6 and GPT-6.1 Sol 23.0. The outlier is the first token: 147.63 seconds to the first answer token, against a 12.24-second median — about 12 times longer — after which output runs at 54 tokens per second, which is unremarkable. OrcaRouter reads this as a queueing and batching story rather than a compute story: a 2.4T sparse model behind a shared endpoint makes requests wait for a slot. And with Alibaba the only provider serving the model, there is no second deployment to separate architecture from serving. Vals AI's own suite points the same way with different arithmetic: 64 minutes 39 seconds and $9.33 per test — a different task mix, so not directly comparable with AA's 32.1 minutes.

Vals Index 63.20%: the top Chinese model

A second independent index tells the opposite story. On the Vals AI Index, Qwen4-Max scores 63.20%, 11th of 38 models — the highest-scoring Chinese entry it tracks, ahead of Kimi K3.1 Pro at 62.52% and Qwen3.8-Max at 59.05%. Read against AA's 43, the point is not which index is right but what each one weights: Vals leans toward white-collar, finance and legal work; AA leans toward terminal work and frontier knowledge. A model tuned for enterprise work looks exactly like this — first on the vendor's own chart, first among Chinese models on Vals, 16 points down on a harness that grades terminal work. OrcaRouter's judgement is that the split itself is the finding: "how good" depends on what you measure.

GA spec corrections: the output ceiling shrinks by a third

The Model Studio product page for the GA model is far more concrete than the preview page, and several figures belong in any selection table. Maximum input is 991,808 tokens without thinking and 983,616 with it; maximum output is 131,072 without thinking, and with thinking on, a 262,144-token budget is shared between reasoning and the answer — a third smaller than the preview's 393,216 shared ceiling. Workloads migrating from the preview that leaned on the larger ceiling lose a third of it. The default reasoning effort moved from xhigh in the preview to high (a 131,072-token thinking budget), so part of the token savings in the official figures may come from running a notch lower out of the box rather than from the architecture. Prices vary by region: Singapore is $1.50/$4.50, while Beijing, Hong Kong, Frankfurt, Virginia and Tokyo carry $1.413/$4.24. There is a 90-day allowance of one million free tokens. Batch inference, fine-tuning and model deployment are all marked unsupported; there are no public weights.

The leak story does not hold: the vLLM patch described a different model

The vLLM pull request #53909, opened August 26 and closed unmerged on October 14, described Qwen3.8-Flash-Next (a 125B-class model), not Qwen4-Max. The three operator names — HyperConnection, QSA and PLE — did survive into the GA model's own documentation under the same names, but the size inference was off by a factor of nineteen on total parameters and twelve on active parameters. The one figure that survives is a ratio: Flash-Next puts QSA on twelve of forty-eight layers, and the flagship's "one layer in four" is the same three-to-one arrangement at whatever depth the larger backbone runs. The leak file got the operators right and the scale wrong.

Checking the "distillation crackdown" claim: a mechanism, no verdict

The other claim circulating that day came from a November 17 post by @kimmonismus: Anthropic's work against distillation is "genuinely hindering the development of Chinese models". OrcaRouter tests it against Anthropic's own three disclosures. The February report described roughly 24,000 fraudulent accounts generating more than 16 million exchanges, naming DeepSeek, Moonshot and MiniMax — not Alibaba. The September threat report attributed the largest campaign it had measured to Alibaba's Qwen lab: a fixed prompt that got Claude to write out its reasoning traces, processed into fine-tuning data for Qwen 3.5, 3.6 and 3.7, totalling over 151 million exchanges between May and July; that channel closed in July. On October 13, Anthropic said its countermeasures had cut extracted reasoning tokens by 90 to 99 percent, naming Zhipu and Moonshot. The verdict is narrow: Qwen4-Max is the generation after 3.5, 3.6 and 3.7, so a channel that closed in July cannot be the reason it underperforms — unless the argument is that the channel carried capability that should have carried forward, a different and much harder claim nobody has evidenced. Kimi is the counter-example: the most-documented exposure, yet it shipped the strongest Chinese model on the board at AA 46. Task mix (Vals weighting white-collar work, AA weighting terminal work and frontier knowledge) explains the two-index split without any outside cause. OrcaRouter does not upgrade the claim to a finding: it has a plausible mechanism, one pattern that fits it, and one counter-example that does not.

A note on effort levels: the 43 still has no ladder

Two points of attribution remain for anyone citing the numbers. First, AA still publishes no effort ladder for this model: "43 is the top rung (xhigh)" is coverage's attribution, not the tracker's. If the 43 turned out to come from the high setting, the gap against Opus would be nine points, not sixteen. Second, the "No publicly available API" placeholder still sits under three charts on the AA model page — a leftover from before the model had a public endpoint, not a finding. The measurement exists; the chart has not caught up.

Our take: the day-two picture is fuller than launch day's and harder to summarise in one line: first among Chinese models on Vals, 16 points behind on AA, a 147-second first token, and a GA output ceiling a third smaller than the preview's. The cheapness is real ($3.27 a task), but if a person waits on the answer, a 147-second first token is a hard blocker; if the work runs unattended as batch agent jobs, 32.1 minutes at $3.27 is a cost and a queue position rather than a failure. Measure the first token on your own traffic before routing.

Reference sources: this article is an original rewrite produced by this site; facts and figures were checked against OrcaRouter — Qwen4-Max at day two: latency lands, leak story does not (Nov 17, 2026) and Artificial Analysis — Qwen4 Max model page (AA Index 43, v5.0.1)

← Back to the Qwen4-Max topic page