Qwen 3.8 Max: Alibaba's 2.4T Parameter Flagship With 16-Day Autonomous Coding
Alibaba just dropped their biggest model ever. Qwen 3.8 Max has 2.4 trillion parameters with 95 billion active, a 1 million token context window, and multimodal capabilities. It ranks #5 on Text Arena, #2 on Vision Arena, and #4 on Frontend Code Arena. Open weights are coming next week.
This is the model that puts Alibaba back in the frontier conversation.
Core specifications
| Spec | Value |
|---|---|
| Total parameters | 2.4T |
| Active parameters | 95B (Sparse MoE) |
| Context window | 1,000,000 tokens |
| Architecture | Sparse Mixture-of-Experts + hybrid attention |
| Multimodal | Yes (text + vision) |
| Release date | August 3, 2026 |
| Open weights | Next week (announced) |
| API | Alibaba Cloud Model Studio |
The 2.4T total with 95B active is a significant jump from Qwen 3.7 Max. The Sparse MoE architecture means the model activates only a fraction of its total parameters per request, keeping inference costs reasonable despite the massive total size.
Benchmark performance
Qwen 3.8 Max does not just compete with frontier models. It beats several of them.
Arena rankings (vendor-reported via Alibaba’s own submissions):
- Text Arena: #5 overall
- Vision Arena: #2 overall
- Frontend Code Arena: #4 overall
Autonomous coding:
- 16-day autonomous coding capability (built “oh-my-cli” from scratch)
- Outperformed human participants in WWW2025 Multimodal Dialogue Intent Recognition Challenge
- Reproduced published research experiments and engineered novel methodologies that outperformed the original papers
Vision capabilities:
- Processes hundred-page documents, full TV series, or 100-hour livestreams
- Converts static inputs into searchable, interactive knowledge bases
- Reconstructs frontend web projects from a single UI screenshot
- Transforms 2D floor plans into 3D interior visualizations
Independent benchmark results (August 2026)
Since the initial launch, third-party evaluation platforms have published independently measured results:
| Benchmark | Score | Source | Context |
|---|---|---|---|
| Terminal-Bench 2.1 | 86.6% | letsdatascience.com | Ahead of Opus 4.8 and Fable 5 (84.6%) |
| Artificial Analysis Intelligence Index | 55.4 | officechai.com | 10 points above Qwen 3.7 Max (46) |
| Artificial Analysis Agentic Index | Ahead of Opus 5 and GPT-5.6 | gurufocus.com | Highest agentic score from a Chinese lab |
Important context on these results:
These are independently measured by Artificial Analysis and other third-party platforms, not self-reported by Alibaba. They represent specific evaluation setups and do not guarantee equivalent performance on every coding or reasoning task.
The Terminal-Bench 2.1 score (86.6%) is particularly notable because it places Qwen 3.8 Max close to GPT-5.6 Sol (88.8%), which was previously the clear leader. However, kingy.ai’s analysis notes that “the evidence supports calling Qwen3.8-Max a credible frontier-class hosted model” while cautioning that it is “not yet proven as a replacement for the strongest coding agents” across all evaluation setups.
The Agentic Index ranking ahead of Opus 5 and GPT-5.6 specifically measures multi-step agent workflows. It does not mean Qwen 3.8 Max is universally superior. Different benchmarks measure different capabilities, and model selection should be based on your specific workload.
Pricing
Updated August 7, 2026: API pricing is now confirmed and live.
| Tier | Input/1M tokens | Output/1M tokens |
|---|---|---|
| Qwen 3.8 Max | $2.00 | $6.00 |
| Qwen 3.7 Max (predecessor) | $2.50 | $7.50 |
At $2/$6 per million tokens, Qwen 3.8 Max is cheaper than its predecessor and significantly undercuts Western frontier models: Claude Opus 5 ($5/$25), GPT-5.6 Sol ($5/$30), and Kimi K3 ($3/$15). It is the cheapest model in the frontier tier by list price, though token consumption per task varies by model.
For the latest pricing across all providers, see our AI API pricing comparison.
Architecture: Sparse MoE with hybrid attention
Qwen 3.8 Max builds on Qwen 3.5’s architecture with two key innovations:
Sparse Mixture-of-Experts: 2.4T total parameters but only 95B active per request. This means the model has massive knowledge capacity while keeping inference costs comparable to much smaller dense models.
Hybrid attention mechanism: Balances massive scale with inference efficiency. The combination of sparse MoE and hybrid attention achieves frontier-level intelligence without the computational overhead of dense models at similar scale.
This architecture is what enables the 16-day autonomous coding capability. The model can sustain long-running tasks without burning through compute budgets.
Multimodal capabilities
Qwen 3.8 Max is not just a text model. It handles vision natively:
Document processing: Hundred-page documents processed in a single pass. Tables, figures, equations, and text extracted with structure preservation.
Video understanding: Full TV series and 100-hour livestreams converted into searchable knowledge bases. The model understands temporal relationships and visual context.
Visual coding: Reconstructs frontend web projects from a single UI screenshot. Transforms 2D floor plans into 3D visualizations. Builds interactive games from natural language descriptions.
Image editing: Edits raw personal footage into professional vlogs. Generates educational animations from text prompts.
Autonomous coding: 16 days without human intervention
The headline capability. Qwen 3.8 Max autonomously executed a real-world software engineering project over 16 days:
- Created “oh-my-cli”, a self-evolving agent framework from scratch
- Established an engineering loop: user feedback, community practices, self-test data
- Iterated through code generation, testing, previewing, and log analysis
- Produced a fully open-sourced framework on GitHub
This is not a benchmark number. This is a real project, running for 16 days, producing real code that works. The model synthesized user feedback and community best practices without human intervention.
For comparison, GLM-5.1 claims 8-hour autonomous coding. Qwen 3.8 Max ran for 16 days. That is a 48x improvement in sustained autonomous operation.
Open weights
Alibaba has promised open weights for the week of August 10, 2026 on Hugging Face and ModelScope. As of August 7, they are not yet downloadable. This will be the first time a Max-class Qwen model goes open-weight.
When released, Qwen 3.8 Max will be the largest open-weight model available, surpassing Kimi K3’s 2.8T (but with only 95B active vs K3’s ~200B active).
Self-hosting a 2.4T model requires serious hardware. Expect to need multiple 80GB+ GPUs or a cloud GPU cluster. For most developers, the API ($2/$6) will be the practical option until community quantization efforts produce smaller variants.
How to access Qwen 3.8 Max
API access
Available now via Alibaba Cloud Model Studio:
- Create an Alibaba Cloud account
- Navigate to Model Studio
- Select Qwen 3.8 Max
- Generate API key
- Start making requests
QwenWork
Alibaba also launched QwenWork, an all-in-one workplace AI agent platform. Qwen 3.8 Max powers the platform’s coding, research, and document processing capabilities.
When to use Qwen 3.8 Max
- You need frontier-class intelligence at mid-tier pricing
- You are building autonomous coding agents that need to run for days
- You process multimodal content (documents, video, images)
- You want open weights for self-hosting (available next week)
- You need 1M context for large codebases or document collections
When to use something else
- Budget coding: GPT-5.6 Luna ($0.20/$1.20) or DeepSeek V4 Flash ($0.14/$0.28)
- Maximum Terminal-Bench: GPT-5.6 Sol ($5/$30, 88.8%) still leads, but the gap to Qwen 3.8 Max (86.6%) is now only 2.2 points
- Best SWE-bench: DeepSeek V4 Pro ($2.19/$8.76, 80.6%)
- Effort level control: Claude Opus 5 ($5/$25, 5-level effort)
- Self-hosting today: Kimi K3 (open weights already available)
My take
Qwen 3.8 Max is Alibaba’s strongest model ever and a legitimate frontier contender. The 2.4T parameters with 95B active is impressive engineering. The 16-day autonomous coding demonstration is the most compelling evidence of sustained agent capability we have seen.
The Vision Arena #2 ranking is particularly notable. Most frontier models focus on text benchmarks. Qwen 3.8 Max competes at the top of both text and vision, which matters for developers building multimodal applications.
The open weights (next week) will be the real test. If the community can run this model efficiently on consumer hardware (through quantization), it could become the default open-weight frontier model. Until then, the API is the practical option.
For developers choosing between frontier models, the decision now is: Qwen 3.8 Max for multimodal and autonomous coding, Kimi K3 for open-weight flexibility, DeepSeek V4 Pro for SWE-bench, and Claude Opus 5 for effort control. Each has a clear niche.
FAQ
How does Qwen 3.8 Max compare to Qwen 3.7 Max?
Qwen 3.8 Max is a significant upgrade. 2.4T params vs undisclosed (estimated 300B+), 95B active vs undisclosed, multimodal support, and 16-day autonomous coding. If you are using Qwen 3.7 Max, upgrade when 3.8 Max API is available.
Is Qwen 3.8 Max open source?
Open weights are promised for the week of August 10, 2026 on Hugging Face and ModelScope. As of August 7, they are not yet available. This is the first Max-class Qwen model to go open-weight. Until weights ship, access is API-only through Alibaba Cloud Model Studio.
How much does Qwen 3.8 Max cost?
$2 per million input tokens and $6 per million output tokens, confirmed via Alibaba Cloud Model Studio. This makes it the cheapest frontier-tier model by list price, undercutting Kimi K3 ($3/$15) and Claude Opus 5 ($5/$25). Note that per-task cost depends on token consumption, which varies by model.
Can Qwen 3.8 Max run locally?
Not yet. Weights are expected the week of August 10, 2026. When released, self-hosting will require multiple 80GB+ GPUs due to the 2.4T total parameter count. Community quantization efforts will likely produce smaller variants for consumer hardware. Until then, use the API at $2/$6.
What is the autonomous coding capability?
Qwen 3.8 Max autonomously built “oh-my-cli”, a self-evolving agent framework, over 16 days without human intervention. It established an engineering loop with user feedback, community practices, and self-test data, iterating through code generation, testing, and log analysis.
How does it compare to Kimi K3?
Both are frontier-class Chinese models. Qwen 3.8 Max has 2.4T/95B active (multimodal, Vision Arena #2). Kimi K3 has 2.8T/~200B active (open weights available now). Qwen 3.8 Max is stronger on vision and autonomous coding. Kimi K3 has open weights available today.