Qwen4-Max's release note mentions open source exactly once: small and medium-sized models in the Qwen4 family will be released gradually.
No sizes. No dates. No licence.
Those 3 blanks say more about Alibaba Group's (9988.HK) thinking than the 2.4 trillion figure does.
Three numbers on the price list
Bailian's Beijing region charges 10 yuan input, 30 yuan output and 0.5 yuan for cache hits, per million tokens. The international site lists $1.5, $4.5 and $0.075. Cache creation is extra: 12.5 yuan for 5 minutes, 20 yuan for an hour. The Qwen app platform matches, with no difference between entrances. At 7.1 to the dollar, $1.5 is 10.65 yuan, about 6% above Beijing.
Cache hits cost one-twentieth of input.
In agent workloads, system prompts, tool definitions and repo indexes are resent every turn, all at the cache rate: a 100,000-token context costs 1 yuan once, then 0.05. Much of the promised "60% lower per-task cost" is billing structure, not a smarter model. The English release calls it "particularly large reductions in cached-input pricing."
A 5-minute cache on 100,000 tokens costs 1.25 yuan, then 0.05 per hit. Two calls pay it back; a single call loses 0.25 yuan.
Caching is not a discount. It is a season ticket for repeated context.
151K tokens per task
On Terminal-Bench 4.0, average tokens per attempt fell from 289K to 151K.
That is 47.8%. The English blog says "about 48%," the Chinese release "near 50%" — three phrasings, one number. Score rose from 37.9 to 48.5, or 28.0%, at the same reasoning effort, Xhigh against Xhigh.
Some coverage used 38.9 for the previous generation, yielding 24.7%; this article uses 37.9, a 3.3-point spread.
Half the tokens, thirty percent more score. Parameter count affects deployment cost; token count affects every invoice. Saving 138K tokens per task across millions of daily calls is a different business.
For two years the race was whose brain was bigger; this year it is who thinks cheaper. On the same chart, Work360 cuts tokens by 24% and 35% at two reasoning tiers, Terminal-Bench by 48% and 49%. Longer tasks save more.
Work360's cost comes in two versions: 26.58 to 11.16 yuan per task in Chinese, down 58.0%, and $4.43 to $1.67 in English, down 62.3%. At 7.1, 26.58 yuan is $3.74, well short of $4.43, so the gap is not exchange rates but likely the tier measured. Four points apart, same direction.
Sixth place, 76 points back
On CodeArena's web development board, Qwen4-Max ranks 6th worldwide.
Alibaba published the rank, not the score. Third-party tabulation puts it at 1750 against 1826 for Claude Fable 5.2, 4.2% behind, on a board voted blind by humans; a footnote credits Arena.
Leaderboards carry folklore. 4.2% does not.
The 328B that goes missing
2.4 trillion is the backbone. The English blog says 2.4T backbone parameters plus 328B of N-gram embeddings, 2.73 trillion combined. Chinese releases drop the embedding. Alibaba leaves the 328B out of the 71B active count, but it still occupies VRAM and still has to be downloaded.
Three changes stand out. Attention is GDN plus QSA: 3 of every 4 layers use Gated DeltaNet to compress history into a fixed-size state, the fourth a lightweight indexer to pick blocks. Residuals become 4 parallel branches with dynamic read-write gating. The vocabulary layer does N-gram lookup at layer 2, buying capacity for little extra compute; GDN plus QSA lowers attention and indexing cost on long sequences.
Training 2.4 trillion parameters produced no loss spikes. And Qwen4-Max-Base scores 5% to 13% lower BPB than Qwen3.8-Max-Base on knowledge, code, QA and STEM, with 25% fewer active parameters.
Lower BPB is better; fewer active parameters are cheaper.
Why the strongest one stays closed
Alibaba has done this before. Qwen3-Max in September 2025 was closed, API and Qwen app only; the largest open release was Qwen3-235B-A22B under Apache 2.0. Qwen3.8-Max in 2026 was closed as well. The top layer earns, the layer below feeds the ecosystem.
The tiering arithmetic runs on active parameters. Open Qwen3-235B activates 22B per pass; closed Qwen4-Max activates 71B, 3.2 times more, so the open model needs under a third of the compute per token. Releasing 2.4 trillion would hand everyone an inference load Alibaba itself finds expensive, and publish its unit-cost curve.
It stays closed because few people could run it anyway.
Size is not the main reason. The same week, Meituan (3690.HK) released LongCat-2.5 at 1.6 trillion under MIT, with a deployment note for 194 Ascend cards. If 1.6 trillion can ship, 2.4 trillion is not too big to open. The difference is the business model: Meituan sells delivery, so the model is a cost line; Alibaba sells cloud, so it is a revenue line. Two decisions, two income statements.
Open source builds ecosystem; closed source builds margin.
What "small and medium" hides
"Small and medium-sized" is where the information hides.
The Qwen3 open list had 8 models from 0.6B to 235B, all Apache 2.0, a 390-fold spread. Against that list, Qwen4 is missing size, date and licence, all 3 still blank.
The list is empty. It will have to write itself.
The biggest beneficiary of open small models is not developers; it is Alibaba Cloud. A small model runs locally, and when it hits a task it cannot do, the user upgrades rather than switches vendors; a large model stays closed, so anyone who needs it calls the API.
One thing is settled today: the 2.4 trillion tier is not on the list.
The open/closed line keeps moving up: 235B in 2025, 1.6 trillion in 2026. Each generation releases the previous flagship and locks the current one, so the community always gets last generation's best, half a step short.
Bad news for people who use models. A moat for people who build them.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
