China’s AI model competition is increasingly being fought on economics as well as benchmark scores. Alibaba’s Qwen3.8-Max pushes scale with 2.4 trillion parameters, while DeepSeek’s smaller V4-Flash attracts attention with unusually low hosted-inference prices. Moonshot AI’s Kimi K3 provides a third approach, pairing very large scale with strong agentic benchmark results.
The comparison shows why a price per million tokens is only the beginning of a purchasing decision. Sparse architecture, active parameters, cache reuse, output length and the number of model calls can all change the cost of completing real work.
Sparse models separate total size from running cost
Qwen3.8-Max uses a mixture-of-experts architecture. Although the model contains 2.4 trillion parameters, Alibaba says around 95 billion are active for each request. DeepSeek V4-Flash contains 284 billion parameters with 13 billion active during inference, according to Artificial Analysis. Kimi K3 has 2.8 trillion total and roughly 104 billion active.
Activating only part of a model can reduce computation compared with using every parameter for every token. It does not, by itself, establish which model is more efficient. Routing design, hardware utilisation, memory movement and the tokens required to solve a task remain important.
Alibaba says Qwen3.8-Max accepts text, images and video, supports a one-million-token context window and completed a software-engineering project lasting 16 days. It reached the leading position among Chinese text models on the crowdsourced Arena.AI ranking after release, while remaining behind several Anthropic models overall. Those results indicate user preference in a particular comparison setting rather than universal superiority.
Alibaba prices Qwen3.8-Max at $2 per million input tokens and $6 per million output tokens. Kimi K3 is listed at $3 and $15 respectively, giving Qwen a clear headline advantage between the two hosted offers.
DeepSeek makes the per-task calculation harder to ignore
Artificial Analysis lists DeepSeek V4-Flash at $0.14 per million input tokens and $0.28 per million output tokens. Its Max Effort variant reportedly offers cache-hit input at $0.003 per million tokens. Cached pricing matters for applications that repeatedly reuse a large body of context, but the saving depends on cache rules and actual hit rates.
In the research firm’s testing, V4-Flash averaged about three cents per test, compared with 86 cents for Kimi K3, $1.86 for OpenAI’s GPT-5.6 Sol and $3.15 for Anthropic’s Claude Fable 5, as reported in the source. DeepSeek’s Max Effort version scored 40 on the firm’s Intelligence Index and generated about 118 tokens per second.
These numbers come from one evaluator and should be interpreted within its workloads and settings. A low task cost is useful only if the result meets the required quality. A model that needs retries, longer prompts or more human correction can erase an apparent API saving.
Kimi K3 illustrates the other side. On the AA-Briefcase agentic knowledge-work benchmark, it reportedly averaged $10.57 per task while producing around 120,000 output tokens across 83 turns. It also achieved the second-highest score in that evaluation at the time. Capability and consumption rose together.
Open weights widen the deployment decision
Alibaba, DeepSeek and Moonshot continue to offer open-weight models alongside hosted APIs. Artificial Analysis lists V4-Flash under an MIT licence, while Kimi K3 uses Moonshot’s own licence. Open weights allow organisations to use their own infrastructure or a third-party provider rather than depend on one hosted endpoint.
That freedom is not free inference. Buyers must account for accelerators, engineering, scaling, observability, security and idle capacity. They also need to inspect each licence and confirm that the available weights match the evaluated version.
A credible model trial should measure complete workflows: successful task rate, tokens consumed, turns, latency, retries and human review. It should include expected context reuse and the organisation’s own output-quality threshold.
Chinese developers are making “good enough, affordable and accessible” a powerful competitive position. The winner for an enterprise will not necessarily be the largest model or the cheapest token. It will be the system that delivers an acceptable result at the lowest repeatable total cost under the organisation’s deployment constraints.
