Open-Weight Frontier
A Chinese flagship just hit 2.4 trillion parameters
Alibaba released its flagship open-weight model, Qwen3.8-Max, setting a new scale record at 2.4 trillion total parameters. The same week, DeepSeek sharpened not its model but its agent execution layer. Since June's Kimi K3 controversy made trillion-parameter models the norm among Chinese labs, the race has expanded beyond raw scale into verifiability and execution reliability.
Alibaba releases Qwen3.8-Max,
its largest open-weight model yet
2.4 trillion total parameters, released same-day under Apache 2.0.
On August 3, Alibaba Cloud announced Qwen3.8-Max, the latest flagship open-weight model from its Qwen team, in an official blog post titled "Introducing Qwen3.8-Max", and released the model weights the same day under an Apache 2.0 license on Hugging Face and its own ModelScope. The model is billed at 2.4 trillion total parameters, currently the largest publicly released open-weight model. Its architecture is a Mixture-of-Experts (MoE) design, activating roughly 18.8 billion parameters per token — about 7.8% of the total. Context length has been extended to 256K tokens.
On the benchmark side, its score on the coding evaluation LiveCodeBench came in at 79.6%, up 11.3 points from the prior-generation Qwen3-Max's 68.3% six months earlier. API pricing is listed at $1.20 per million input tokens and $4.80 per million output tokens on Alibaba Cloud Model Studio's rate card. Zhou Wenbo, a senior research lead on the Qwen team, said in the blog post that the team was "committed to shipping the largest model that can run on open weights, at a price that's actually usable in production."
June's Kimi K3 controversy pushed
trillion-parameter scale into the norm
Distrust in self-reported benchmarks turned into pressure for verifiability.
Understanding this release requires going back to June's "Kimi K3 controversy." Moonshot AI's Kimi K3, a 1.2-trillion-parameter open-weight model released in June, saw its vendor-reported LiveCodeBench score come in nearly 9 points lower under independent reproduction, as flagged in an Artificial Analysis review. That episode sparked industry-wide distrust of self-reported numbers. Since then, being "trillion-parameter" has become table stakes among top-tier Chinese models, and labs have shifted toward releasing weights same-day and structuring announcements so third parties can verify them.
Qwen3.8-Max's rollout — an official blog post, weight release, and benchmark report, rather than a press release leading the news — is likely not unrelated to that context.
| Kimi K3 (June, Moonshot AI) | Qwen3.8-Max (Aug, Alibaba) |
|---|---|
| 1.2T total parameters | 2.4T total parameters |
| Vendor score disputed on reproduction | Benchmark report and weights released together |
| Limited license terms | Apache 2.0, released same day |
The release, by the numbers
The same week, DeepSeek V4
sharpened its agent harness
Differentiating not on model size, but on how reliably it uses tools.
Worth watching alongside Qwen3.8-Max is DeepSeek's move. DeepSeek V4 itself, a 1.8-trillion-parameter model released in May, hasn't changed much — but on August 4, DeepSeek published a new orchestration layer, "DeepSeek Agent Runtime 2.0," designed to improve tool-call reliability, on GitHub, raising its score on τ-bench, an agentic-task evaluation, from 61.4% to 74.2%. Rather than competing on model parameter count, this is differentiation at the execution layer — getting an existing model to use tools correctly. Li Hao, the principal engineer leading DeepSeek's agent infrastructure work, wrote in the release notes that "most tool-call failures aren't a reasoning shortfall — they come down to how execution order and retry logic are designed."
A trillion parameters no longer turns heads.
The next axis of competition is verifiability and execution.
Why this matters now
Through 2024–2025, the Chinese open-weight race was mostly about closing the performance gap with GPT- and Claude-class models. Since June's Kimi K3 controversy, the contest has broadened to cover both "parameter scale" and "verifiability." Qwen3.8-Max set a new record on total parameters, while DeepSeek, in the same window, chose to strengthen its agent execution layer rather than the model itself — a sign that the pure scale race may be settling down as competition shifts toward reliability in production use. That the two labs moved in different directions in nearly the same week is itself evidence of how broad the industry's focus has become.
Who it affects, and how
If you're evaluating models in-house as an engineer, Qwen3.8-Max and DeepSeek V4 now belong on the same shortlist. The former stands out on raw model performance and context length; the latter on tool-call success rate — and if your workloads are agentic, DeepSeek's execution-layer improvements shouldn't be dismissed. For PMs and product leads, this episode makes clearer than ever the risk of choosing a model purely on benchmark score — it's worth building a comparison sheet that also accounts for third-party verification and license terms. For leadership, while not a direct decision point, the fact that Chinese open-weight models are approaching production-grade reliability is worth keeping in mind as an option for avoiding vendor lock-in.
What happens next, and recommended actions
In the near term, watch whether Qwen3.8-Max's API expands to overseas regions, and whether DeepSeek Agent Runtime 2.0 can be paired with other open-weight models. (1) Evaluation teams should run both models against benchmarks close to their own workloads — LiveCodeBench, τ-bench, or similar — and check the gap against vendor-reported numbers themselves. (2) Teams running agentic workflows should consider improving the orchestration layer, along the lines of DeepSeek Agent Runtime 2.0, before swapping out the underlying model. (3) Given the Kimi K3 episode, keep an eye on whether future announcements from Chinese labs pair released weights with a way to verify them.
Counter-view, risks, and limits
The 2.4 trillion figure is a total, easy to overstate given only about 7.8% of it actually activates at a time. A single-benchmark gain like LiveCodeBench also doesn't guarantee real-world usefulness — and as the Kimi K3 episode showed, vendor-reported numbers should always be checked against independent verification. DeepSeek Agent Runtime 2.0's τ-bench improvement is likewise fresh off release, with no independent reproduction yet available, so for now it remains a primary-source claim rather than a verified one. Both models also carry constraints tied to China's domestic regulatory and data-governance requirements, so usage terms should be checked individually depending on region.