Alibaba Launches 2.4-Trillion-Parameter Qwen3.8-Max as DeepSeek Slashes AI Inference Costs
Alibaba has introduced Qwen3.8-Max, its largest artificial intelligence model to date, while rival Chinese developer DeepSeek is gaining attention for offering one of the lowest inference prices among major AI systems.
The two releases highlight the intensifying competition among Chinese AI companies, not only in model performance and scale, but also in affordability, openness, and real-world deployment costs.
Alibaba Unveils Its Largest AI Model
Qwen3.8-Max contains 2.4 trillion parameters and uses a mixture-of-experts, or MoE, architecture.
Instead of activating the entire model for every request, the system uses only the portions needed to process a particular task. Alibaba said approximately 95 billion parameters are active during each inference request.
This design can reduce computing costs and response delays compared with models that activate all their parameters simultaneously.
DeepSeek uses a similar sparse architecture at a smaller scale. Artificial Analysis lists DeepSeek V4-Flash at 284 billion total parameters, with only 13 billion activated during inference.
Moonshot AI’s Kimi K3, meanwhile, contains 2.8 trillion total parameters and activates approximately 104 billion at a time.
Qwen3.8-Max Handles Text, Images, and Video
Qwen3.8-Max is a multimodal model capable of processing text, images, and video.
It supports a context window of up to one million tokens, allowing it to analyse long documents, extensive conversations, large codebases, and complex multimedia content within a single session.
Alibaba also said the model successfully completed a software engineering project that lasted 16 days, suggesting that the system is designed for long-running and highly complex agentic workflows.
In terms of size, Qwen3.8-Max is close to Moonshot AI’s Kimi K3, which was released in July.
The two companies are also competing aggressively on API pricing.
Alibaba charges:
- $2 per million input tokens
- $6 per million output tokens
Kimi K3 costs:
- $3 per million input tokens
- $15 per million output tokens
However, model size and advertised token prices do not fully determine how expensive an AI system is to operate.
Architecture, active parameter count, output length, token consumption, and the number of model calls required to complete a task can all affect the final cost.
Qwen3.8-Max Climbs Chinese AI Rankings
Following its release, Qwen3.8-Max reportedly moved into the top position among Chinese text models on the crowdsourced comparison platform Arena.AI.
The model remained behind several Anthropic systems in the overall rankings.
It also placed second on Arena.AI’s leaderboard for models that analyse images and other visual content, ranking behind an Anthropic Claude Fable 5 variant.
These rankings suggest that Chinese models are closing the performance gap with leading Western AI systems while maintaining a strong focus on lower pricing and open deployment options.
DeepSeek Pushes AI Inference Prices Lower
DeepSeek has taken a different approach with V4-Flash.
Instead of trying to match the enormous size of Alibaba’s Qwen3.8-Max or Moonshot AI’s Kimi K3, DeepSeek has focused on lowering the cost of using advanced AI.
According to Artificial Analysis, DeepSeek V4-Flash costs:
- $0.14 per million input tokens
- $0.28 per million output tokens
The model reportedly supports a one-million-token context window and contains 284 billion total parameters, with 13 billion activated during inference.
Artificial Analysis also lists cache-hit pricing of only $0.003 per million tokens for the Max Effort version of V4-Flash.
That rate is approximately 98 percent lower than the model’s standard input price.
Cached input refers to information that the model has already processed and can reuse in later requests. This can significantly reduce costs for applications that repeatedly work with the same documents, instructions, databases, or conversation histories.
DeepSeek Records Lower Cost Per Benchmark
DeepSeek’s lower token pricing was also reflected in benchmark testing conducted by Artificial Analysis.
Reuters reported that the research firm estimated DeepSeek V4-Flash’s average cost at approximately three cents per test.
By comparison, the estimated benchmark costs were:
- Kimi K3: $0.86 per test
- OpenAI GPT-5.6 Sol: $1.86 per test
- Anthropic Claude Fable 5: $3.15 per test
The comparison considers the amount of input and output each model uses while completing the benchmark.
This distinction is important because a low per-token price does not automatically guarantee a lower total cost.
A model may still become more expensive if it generates extremely long responses, makes repeated tool calls, or requires multiple interactions before reaching the correct answer.
Artificial Analysis gave the Max Effort reasoning version of DeepSeek V4-Flash a score of 40 on its Intelligence Index.
The research firm also measured an output speed of approximately 118 tokens per second during testing.
Token Prices Do Not Tell the Whole Story
Moonshot AI’s Kimi K3 illustrates how advertised API prices can differ significantly from the actual cost of completing complex workloads.
Artificial Analysis lists Kimi K3 at:
- $3 per million input tokens
- $15 per million output tokens
- $0.30 per million cached input tokens
On the Artificial Analysis AA-Briefcase benchmark for agentic knowledge work, Kimi K3 reportedly cost an average of $10.57 per completed task.
The model generated approximately 120,000 output tokens and required an average of 83 turns for each task.
Artificial Analysis said the total cost was influenced by Kimi K3’s token pricing, output volume, and number of model interactions.
This means that repeated calls and lengthy responses can push the actual cost of an AI workload far beyond the advertised API rate.
Despite the higher task cost, Kimi K3 recorded the second-highest overall score on the AA-Briefcase evaluation at the time of testing, ranking behind Claude Fable 5.
It also received a score of 57 on Artificial Analysis’ broader Intelligence Index.
The comparison between Kimi K3 and DeepSeek V4-Flash shows why developers and businesses should examine cost per completed task, not just cost per million tokens.
Models with different architectures and reasoning patterns may consume vastly different amounts of compute and tokens while attempting to solve the same problem.
Open-Weight Models Give Developers More Control
Pricing competition is also being shaped by how Chinese AI companies distribute their models.
Alibaba, DeepSeek, and Moonshot AI continue to support open-weight releases alongside hosted API services.
Open-weight models allow developers to download and deploy the underlying model parameters on their own infrastructure or through third-party cloud providers.
Artificial Analysis lists DeepSeek V4-Flash as an open-weight model released under an MIT licence, with its weights available through Hugging Face.
Kimi K3 is also offered as an open-weight model under Moonshot AI’s own licence.
This approach gives developers more flexibility and reduces dependence on a single company’s hosted inference platform.
Running a large AI model independently can still require expensive computing hardware, energy, storage, and technical expertise. However, developers gain greater control over data privacy, customisation, security, and deployment.
This differs from the primary commercial models offered by OpenAI, Anthropic, and Google, which generally keep their model weights closed.
Businesses May Prioritise Affordability Over Maximum Performance
Lian Jye Su, chief analyst at Omdia, said many organisations do not necessarily require the most powerful AI model available.
“Many business workflows do not need the industry’s very best model,” Su said. “They need models that are good enough, affordable, transparent and accessible, and open-weight models help meet that demand.”
For businesses, the emerging competition among Alibaba, DeepSeek, and Moonshot AI offers more choices.
Companies can select massive frontier models for complex reasoning and multimodal workloads, or choose smaller and cheaper models for customer support, document processing, coding assistance, translation, content production, and automated business operations.
The new generation of Chinese AI models shows that the global AI race is no longer focused solely on building the biggest system.
The competition is increasingly about delivering acceptable intelligence at the lowest practical cost, while giving developers greater control over how and where AI models are deployed.
For Filipino developers, startups, educators, creators, and enterprises, this could mean access to more affordable AI technologies capable of handling Filipino languages, local business processes, and large-scale digital services.
The most powerful model may continue to dominate the headlines, but for many organisations, the real winner will be the model that can complete the job reliably without exhausting the budget.
