Alibaba Launches Qwen3.8-Max as DeepSeek's Low Pricing Intensifies China's AI Competition
Taylor Wilson
Alibaba released Qwen3.8-Max with 2.4 trillion parameters on the same day DeepSeek's V4-Flash entered at roughly 1% of Claude Fable 5's cost — China's AI race now spans scale, efficiency, and distribution strategy all at once.
How big is Qwen3.8-Max, really?
Total parameters: 2.4 trillion. Active per inference: 95 billion — it uses a Mixture-of-Experts architecture (MoE, splitting the model into specialist modules and activating only the relevant ones per query).
This means → the model is massive on paper, but actual compute per request is far smaller than running all parameters, cutting both cost and latency.
It accepts text, image, and video inputs with a context window up to 1 million tokens.
Do independent rankings back Alibaba's claims?
Alibaba's internal tests placed the preview at second globally, behind only Claude Fable 5. Independent platform Arena is more conservative: text #5, vision #2, coding #4.
On text, it trails four Anthropic models; on coding, it trails Claude Opus 5 Max, Kimi K3, and Claude Opus 5 High.
In plain terms = it is the highest-ranked Chinese model in Arena's text category, but a gap to the global frontier remains — Alibaba's internal benchmark reads optimistic.
How low is DeepSeek's pricing?
V4-Flash: $0.14 per million uncached input tokens, $0.28 per million output tokens.
Artificial Analysis estimates average cost for equivalent benchmark tasks: V4-Flash ≈ $0.03, Kimi K3 ≈ $0.86, GPT-5.6 Sol ≈ $1.86, Claude Fable 5 ≈ $3.15.
This means → V4-Flash completes the same work at roughly 1% of Claude Fable 5's cost — not "a bit cheaper," but an order-of-magnitude gap.
Where do Alibaba and Moonshot sit on price?
Qwen3.8-Max: input $2/M tokens, output $6/M tokens.
Kimi K3: input $3, output $15 — roughly double Alibaba's rate.
This reflects a three-tier pricing map forming in China's AI market: DeepSeek ultra-low, Alibaba mid-range, Kimi premium — each positioning for a different logic.
What does "open-weight" actually open?
DeepSeek V4-Flash is already released as open-weight; Alibaba plans to follow next week.
In plain terms = open-weight means developers can download and customize trained model parameters, but it does not necessarily include training data or code — it is not full open-source.
Omdia chief analyst Lian Jye Su notes: many workflows don't need the very best model — they need one that is good enough, affordable, transparent, and accessible. Open-weight models fit that demand.
What does the "DeepSeek dead zone" mean?
Within eight weeks, multiple Chinese models have reached or approached global benchmark leaders, creating a so-called "DeepSeek dead zone."
This means → any product that offers less capability at the same price, or the same capability at a higher price than DeepSeek, faces survival pressure.
Qwen3.8-Max and Kimi K3 currently hold the high-performance tier, but whether they can sustain a differentiation gap on cost is the key variable determining if this competitive structure holds.
Content is for reference only, not financial advice.