Alibaba's Qwen3.8-Max nearly triples its autonomous coding score and promises open weights for a 2.4-trillion-parameter model

Alibaba's Qwen3.8-Max nearly triples its autonomous coding score and promises open weights for a 2.4-trillion-parameter model

Alibaba's Qwen team formally released Qwen3.8-Max, a 2.4-trillion-parameter multimodal model that scored 56.6 on the DeepSWE autonomous coding benchmark, up from 21.6 for its predecessor Qwen3.7-Max. 1

That 56.6 score is a 2.6-fold jump in a single model generation, and it places Qwen3.8-Max alongside DeepSeek V4 Flash (54.4) on autonomous coding tasks. 12 The DeepSWE benchmark measures how well AI models autonomously complete real-world software engineering tasks across 113 tasks spanning 91 repositories and 5 programming languages, with all models running on a standardized harness for consistency. 3

The proprietary-to-open gap has collapsed to five percentage points

On the independent DeepSWE leaderboard maintained by DataCurve, the proprietary frontier stands at 74% for Claude Opus 5 and 73% for GPT-5.6 Sol. 3 The highest-scoring open-weight model, Kimi K3 from Moonshot AI, scores 69% with a 5-point margin of error. That puts the gap between the best proprietary model and the best open-weight model at five percentage points, with overlapping confidence intervals (Kimi K3: 64 to 74%, Claude Opus 5: 70 to 78%). On this benchmark, the open-weight frontier is statistically indistinguishable from the proprietary one.

The cost picture mirrors the score picture. Per DeepSWE task, Kimi K3 costs $4.65. Claude Opus 5 costs $11.84 and GPT-5.6 Sol costs $8.39. 3 Divided by score, each percentage point of DeepSWE performance costs $0.067 from the open-weight leader and $0.160 from the proprietary frontier. The best proprietary model charges 2.4 times more per point of coding capability for a five-point gap that sits inside the margin of error.

Here is the caveat that matters for anyone evaluating Qwen3.8-Max: its 56.6 score does not appear on the independent leaderboard. The number comes from Qwen's own benchmark table, published alongside the model's release. 4 Qwen3.8-Max has not been run through DataCurve's standardized harness. If it were, the score could land higher or lower. Even independently verified models show harness-dependent variation: Moonshot AI's own testing put Kimi K3 at 67.5 on DeepSWE, while the leaderboard's standardized harness reports 69%. 5

If Qwen3.8-Max's self-reported 56.6 holds up under independent testing, it would slot between Claude Opus 4.8 (59%) and Claude Sonnet 5 (54%) on the leaderboard: second among open-weight models, behind Kimi K3, but still 17 points behind the proprietary frontier. 3

The open-weight trillion-parameter race has compressed to days

Qwen3.8-Max's announcement on July 19 landed three days after Moonshot AI revealed Kimi K3, a 2.8-trillion-parameter model with open weights planned for July 27. 1 The timeline of Chinese open-weight frontier models now reads as a sprint:

• DeepSeek V4 Flash: open weights available, DeepSWE 54.4 (vendor-reported) 2
• GLM-5.2: open weights available, DeepSWE 44% (independently verified on leaderboard) 3
• Kimi K3: announced July 16, 2.8T parameters, open weights planned July 27, DeepSWE 69% (independently verified) 53

But Qwen's open-weight pledge carries a specific history. Both Qwen3.7-Max and Qwen3.6-Max-Preview shipped as closed models available only through Alibaba Cloud's Model Studio. 6 Qwen3.8-Max would be the first in the Max tier to break that pattern.

Ten days of autonomous development, by Alibaba's account

Qwen says Qwen3.8-Max ran for more than 10 days of autonomous self-evolving development, building production-quality deliverables from an empty folder without human intervention. 7 The QwenCloud product page describes the model delivering complete projects spanning over 10 days across legal, financial, design, and other professional domains, end-to-end in a single conversation. 8

These are vendor claims without independent replication. But they signal where Alibaba is positioning Qwen3.8-Max: not as a code completion tool, but as an autonomous agent that can sustain work over timescales measured in days. Alibaba also self-positions the model as second only to Fable 5 among all available models, a claim no third-party evaluation has confirmed. 19

What this means for builders and investors

The strategic question is no longer whether open-weight models can reach the frontier. On the DeepSWE benchmark, they already have. 3

The question is how fast the rest of the field catches up, and whether Alibaba follows through on open weights for a Max-tier model for the first time. Investors tracking the proprietary AI sector face a harder calculation: when frontier-class coding capability arrives as downloadable weights within weeks of a closed model's release, the competitive moat around proprietary model access narrows to whatever proprietary systems can do that open ones cannot, and the window for that gap is measured in weeks, not quarters.