DeepSeek V4 Goes Live, Alibaba Fires Back With 2.4T-Parameter Qwen 3.8DeepSeek V4正式发布，阿里随即推出2.4万亿参数Qwen 3.8

DeepSeek V4 moved from preview to general availability on July 19th, three months after its April release. The concrete new trigger this cycle: dynamic off-peak pricing set at 87 cents per million tokens for V4-Pro — compared with $50 for the same volume from Anthropic's Fable 5. DeepSeek also introduced on-peak/off-peak tiering, a pricing architecture no Chinese frontier lab had deployed at this scale before.

Within hours, Alibaba's Qwen team announced Qwen 3.8: a 2.4-trillion-parameter open-weight model already live in preview on Alibaba's Token Plan and Qoder coding platforms, priced at one-tenth of standard API rates. Alibaba describes the model as 'second only to Fable 5.' No independent benchmark exists yet for Qwen 3.8; the parameter count and the pricing are the verifiable facts.

The combined move puts API cost pressure on every lab operating in the China market and on global resellers of Western models. The pattern — release open weights, set preview pricing near zero, then raise — has now been executed by both DeepSeek and Alibaba within the same 48-hour window.

Alibaba's Qwen team moved on Sunday with the kind of timing that makes competitive intent unmistakable. Within hours of Moonshot AI's Kimi K3 open-weight debut at WAIC in Shanghai, the Qwen team pushed Qwen3.8-Max-Preview live on its Token Plan and Qoder coding platforms — priced at one-tenth of the standard API rate. The sequence was not coincidental. It was choreography.

The headline number is 2.4 trillion total parameters, which Alibaba positions as second only to "Fable 5" among publicly benchmarked systems. MarkTechPost reported the launch alongside its WAIC context, noting the deliberate proximity to Moonshot's announcement. Cybernews framed it as part of a coordinated cost-pressure campaign from Chinese frontier labs. Silicon Republic picked up the parameter count. All three missed the more important absence.

What is not present matters more than what is. There is no model card. There is no confirmed license — Apache or otherwise. Most critically, there is no active parameter count. For a mixture-of-experts architecture, which 2.4T strongly implies, the active figure is the number that determines real inference cost per token. Without it, the 2.4T headline is a marketing datum, not a technical one. At 4-bit quantization, storing all weights requires roughly 1.2 terabytes — awkward arithmetic for even a dense eight-card H200 node. Whether a typical expert-routing configuration would bring that to something deployable at the claimed price point remains unanswered.

The pricing signal, however, is real and worth reading carefully. One-tenth of standard API rates for a preview model is not a discount — it is a stake in the ground. Alibaba is telling prospective enterprise customers and developers: whatever the capability frontier currently costs, Qwen will cost less. This mirrors the strategy that made Qwen2.5's earlier releases disruptive: release early, price aggressively, let the benchmark card catch up. Watch whether the preview rate survives past the first 30 days, and whether it anchors the post-release GA price.

Community reaction on Hacker News split along a now-familiar fault line: enthusiasm among developers who want another credible open-weight entrant at the frontier, and fatigue from a cohort that has watched vendor benchmark claims outpace reproducible evaluations repeatedly over the past 18 months. That fatigue is legitimate. The Qwen team has published rigorous technical reports in the past — the Qwen2.5 series technical report is detailed and reproducible — which is precisely why the absence of equivalent documentation here is notable rather than ignorable.

The competitive pressure Alibaba is applying lands on multiple targets simultaneously. Domestically, it arrives as a response to Moonshot's open-weight credibility play at a flagship state venue. Internationally, it positions the Qwen family against Meta's Llama and Mistral in the cost-per-capable-token conversation. The 1/10th price, combined with the Qoder integration — Alibaba's coding-assistance product — suggests the initial deployment target is enterprise coding workflow adoption, where switching costs are low enough that price actually converts.

What you should watch: First, the model card publication date. If it arrives within two weeks, the preview framing holds. If it stretches past 30 days, the "preview" label is doing work the documentation should be doing. Second, the active parameter disclosure — this single number will tell you whether the serving economics behind the 1/10th price are structurally sustainable or promotional. Third, watch whether Alibaba confirms or denies an open-weight release on Hugging Face or ModelScope; the absence of a license statement in the preview is legally and commercially consequential for any enterprise considering integration. The pricing is the sentence. The model card is the speech. Right now, you only have one of them.

DeepSeek introduced dynamic pricing alongside V4's general availability on Wednesday: during off-peak hours, V4 Pro runs at $0.87 per million tokens — compared to $50 per million for Anthropic's Fable 5, a 57-to-1 differential. Early testers report V4's performance approaching Claude Opus 4.8 and OpenAI's GPT-5.6 Sol, though no independent benchmark has been published.

Hours after DeepSeek's GA announcement, Alibaba's Qwen team unveiled Qwen 3.8, an open-weight model at 2.4 trillion parameters — which would rank among the largest open-weight releases ever. A preview is live on Alibaba's Token Plan and Qoder platforms at one-tenth of standard API rates. Alibaba describes the model as 'second only to Fable 5,' but again without third-party corroboration.

The pricing structure follows an established pattern: Chinese labs are matching frontier capability claims, then undercutting on cost to drive API adoption ahead of benchmark verification. The combination of V4 and Qwen 3.8 represents the most aggressive dual-pricing push from Chinese labs in a single week.

Alibaba previewed Qwen 3.8-Max on July 19 at WAIC, claiming it is 'second only to Fable 5' among frontier models. As of July 21-22, multiple independent reviewers confirm the same finding: no benchmark table has been published, no model card exists, the active-parameter count is undisclosed, and the weights are not on Hugging Face or any public repository — making the '2.4 trillion parameter' figure a headline number that says little about serving cost or verifiable capability.

The preview is accessible through Alibaba's Token Plan at 10% of standard pricing, and open weights are promised 'soon' — but Alibaba has not named a date or license. Qwen's flagship Max tier has historically remained closed-source, making the open-weight pledge a departure from established practice.

The credibility gap matters beyond marketing: if Alibaba follows through on open weights, Qwen 3.8 would, by parameter count, exceed every open-weight model currently available, including DeepSeek V4 Pro at 1.6 trillion. Without independent evaluation, the claimed ranking against Anthropic's Fable 5 cannot be assessed.