Qwen3.8-Max: 2.4T Parameters, Open Weights, and the First Model to Code Autonomously for 16 Days

Executive Summary

On August 3, 2026, Alibaba's Qwen team officially released Qwen3.8-Max, the most capable model in the Qwen family to date and the first open-weight model to exceed 2 trillion parameters. Built on a sparse Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters and 95 billion activated per token, the model delivers frontier-level performance across coding, professional work, research, and long-horizon autonomous tasks — while committing to release open weights within a week.

What sets Qwen3.8-Max apart is not just scale but demonstrated long-horizon autonomy: the model autonomously built a self-evolving agent framework over 16 continuous days (265 commits, 127 PRs, 151 issues), reproduced a research paper from scratch and then improved upon it over 5 days of self-directed experimentation, and beat 87% of human teams in a live multimodal competition within 24 hours. These are not benchmark scores — they are real-world execution traces with publicly available artifacts.

The model ranks 5th in Text Arena and 2nd in Vision Arena, scoring 86.6 on Terminal-Bench 2.1 (ahead of Claude Opus 4.8 at 84.6), 56.6 on DeepSWE 1.1 (a 2.6× improvement over Qwen3.7-Max), and 73.5 on FrontierSWE. At $2.00/M input tokens and $6.00/M output tokens (with cached reads at $0.25/M), it positions itself between the ultra-cheap DeepSeek V4-Flash and the premium U.S. frontier models, while the upcoming open-weight release threatens to collapse that pricing advantage entirely.

This article provides a comprehensive technical analysis of the Qwen3.8-Max release, its architectural innovations, benchmark performance, long-horizon execution capabilities, and the strategic implications for the global AI landscape.

---

1. The Release: First Open-Weight Max-Class Model

1.1 What Is Qwen3.8-Max?

Qwen3.8-Max represents a generational leap for the Qwen family:

• 2.4 trillion total parameters — the largest open-weight model ever announced
• 95 billion activated parameters per token — approximately 4% of total parameters active during inference
• 1 million token context window — with 991K max input (983K with thinking enabled) and 131K max output
• Multimodal input — accepts text, images, and video; returns text
• Sparse MoE architecture — built on the Qwen 3.5 foundation with hybrid attention
• Open weights promised — first Qwen-Max-class model to be open-sourced, with weights releasing the week of August 10, 2026
• Companion model — Qwen3.8-27B also going open-weight for on-premise deployment

1.2 The Qwen Family Context

Model	Total Params	Activated	Context	Open Weights	Release Date
Qwen3.8-Max	2.4T	95B	1M	Yes (next week)	Aug 3, 2026
Qwen3.8-27B	27B	27B (dense)	1M	Yes (next week)	Aug 3, 2026
Qwen3.7-Max	Not disclosed	Not disclosed	1M	No (API-only)	May 19, 2026
Qwen3.7-Plus	Not disclosed	Not disclosed	1M	No (API-only)	May 2026
Qwen3.5	Not disclosed	Not disclosed	1M	Yes	2026

1.3 Availability

• API: Available now via QwenCloud (OpenAI- and DashScope-compatible)
• Hosted: Qwen Studio (chat.qwen.ai) with the model selectable
• Open weights: Expected the week of August 10, 2026 (license TBA)
• Enterprise: Alibaba Cloud Model Studio for global developers

---

2. Architecture: Sparse MoE at 2.4 Trillion Parameters

2.1 The Scale Decision

At 2.4 trillion total parameters with 95B activated, Qwen3.8-Max represents a bold bet on extreme sparsity. The model activates only ~4% of its total parameters per token, a much higher sparsity ratio than most MoE models in the frontier:

Model	Total Params	Activated	Sparsity
Qwen3.8-Max	2.4T	95B	96% sparse
DeepSeek V4-Flash	284B	13B	95.4% sparse
DeepSeek V4-Pro	1.6T	49B	96.9% sparse
Kimi K3	2.8T	Not disclosed	Not disclosed

2.2 Hybrid Attention Mechanism

Built on the Qwen 3.5 foundation, Qwen3.8-Max features a hybrid attention mechanism that combines:

• Standard self-attention for local context processing
• Compressed attention for long-range dependencies across the 1M-token context window

This design allows the model to maintain frontier-level reasoning while keeping inference costs manageable despite the massive total parameter count.

2.3 Reinforcement Learning at Scale

A key differentiator is the joint scaling of RL environments and compute. Alibaba's approach addresses three coupled challenges:

1. Environment scaling — Continuously expanding decoupled real environments along independent axes: Task (single → multi → multi-day), Workspace (multi-file → hierarchical → complex heterogeneous), and Harness (category, version, skills)
2. Universal Reward System — Internalizing heterogeneous verification (execution-based checking, rubric-conditioned adjudication, agentic inspection) under automatically scalable rubrics
3. Online data balancer — Shaping every training batch to maintain distribution balance across tasks, difficulty, workspaces, and harnesses

The result is a measurable horizontal lift in real-world working ability across dozens of benchmarks as RL training scales.

---

3. Long-Horizon Autonomous Coding: The 16-Day Project

3.1 The oh-my-cli Project

The most compelling demonstration of Qwen3.8-Max's capabilities is the oh-my-cli project — a self-evolving agent framework built entirely autonomously over 16 continuous days:

• 265 commits across the repository
• 127 pull requests created, reviewed, and merged
• 151 issues tracked and resolved
• Zero human intervention throughout the entire process

The project implements a complete engineering loop:

Key implementation details:

• Loop Engineering Setup: Task state machine (ready → leased → active), dispatcher, monitor, and watchdog
• Self-testing: Build, Unit Test, E2E, and Desktop Lifecycle validation after each update
• Multi-source Evolution: Converting community experience and user feedback into executable work

The complete project trace is publicly available at github.com/qwen-code-dev-bot/oh-my-cli.

3.2 Research Paper Reproduction and Improvement

Qwen3.8-Max was given a research paper — "Unified Data Selection for LLM Reasoning" — and asked to reproduce the experiment from scratch, then improve upon it.

The results:

• ~125 hours of continuous autonomous work (~5 days)
• ~7,600 lines of code written from scratch
• 1,100+ actions taken
• 33 rounds of GPU training executed
• 18 improvement ideas tested across 4 rounds

Round	Best Idea	Score (AIME24)	Gain vs. Baseline
—	Paper's method (reproduced)	49.58%	—
1	Split data by difficulty	50.42%	+0.84
2	Weight by entropy–score gap	51.67%	+2.09
3	Tune selection width	51.25%	+1.67
4	Count hard decision points (nhighgate) ★	52.29%	+2.71

The model not only reproduced the paper's six main findings but invented a new method that beat the paper's own approach by 2.7 points on the competition-level math benchmark.

3.3 Live Competition: Beating 87% of Human Teams

Qwen3.8-Max entered the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba Cloud's Tianchi platform, competing against 526 human teams.

• 24-hour time limit, fully autonomous
• 45 submissions with iterative improvement
• Final accuracy: 0.853, beating 458 of 526 teams (87%)
• Approach: Fine-tuned and ensembled BERT, MacBERT, RoBERTa for text; Qwen2.5-VL-7B + Chinese-CLIP for images; weighted-voting fusion with cross-validation calibration

---

4. Benchmark Performance

4.1 Agentic Coding Benchmarks

Benchmark	Qwen3.8-Max	Qwen3.7-Max	Claude Opus 4.8	Claude Fable 5	GPT-5.6 Sol
Terminal-Bench 2.1	86.6	69.7	84.6	84.6	88.8
DeepSWE 1.1	56.6	21.6	~58	~55	~55
FrontierSWE	73.5	40.7	~75	88.8	~70
SWE-bench Pro	67.7	60.6	69.2	80.0	64.6
JobBench	53.4	31.3	—	—	—

4.2 Reasoning & Knowledge Benchmarks

Benchmark	Qwen3.8-Max	Qwen3.7-Max
GPQA Diamond	92.6	92.4
PaperBench	93.0	—
IFBench	82.8	—

4.3 Multimodal Benchmarks

Benchmark	Qwen3.8-Max
OSWorld-Verified	86.1
Parametric CAD Bench	91.5
OmniDocBench 1.5	92.1
RecreationBench	Frontier (black-box app reconstruction)

4.4 Arena Rankings

• Text Arena: #5 of 215 (score: 78.16/100 on BenchLM)
• Vision Arena: #2

4.5 Key Observations

1. Terminal-Bench 2.1 leap: From 69.7 (Qwen3.7-Max) to 86.6 — a 23.8% relative improvement in terminal-based coding
2. DeepSWE 1.1: 2.6× improvement over predecessor (21.6 → 56.6)
3. Vision dominance: Leading scores across OSWorld, CAD, and document benchmarks
4. SWE-bench Pro gap: Still trails Fable 5 (67.7 vs. 80.0), indicating room for improvement in harder software engineering tasks
5. RL scaling curve: Alibaba's own data shows the RL scaling curve peaks at ~0.725 near 4,000 training environments, then declines to 0.719 and 0.689 — suggesting diminishing returns at current scale

---

5. Professional Work Capabilities

5.1 Breadth Across Hundreds of Professions

Qwen3.8-Max was stress-tested across high-value professional workflows:

Profession	Task	Time	Traditional Equivalent
Corporate Counsel	Review 1,284 clauses across hundreds of documents	<1 hour	Paralegal team, ~1 week
UI/UX Designer	8-screen interactive banking app prototype	1 shot, 0 revisions	3-5 revision rounds
Restaurant Founder	26-dish menu with caloric/provenance data	1 pass	Chef + ops team, weeks
Structural Engineer	Seismic model of 30-story tower	Real-time browser	Specialized software, >1 week
Rehab Therapist	3D interactive injury visualization	Minutes	Medical animation studio, 2-4 weeks
Sports Analyst	8,400 possessions → tactical profile	Tens of minutes	Analytics team, several days

5.2 Quantitative Finance: End-to-End Strategy R&D

Qwen3.8-Max built a complete ETF-rotation strategy from a one-line description:

• Depth: Autonomous data system construction, factor building, multi-round greedy iteration, dynamic backtest analysis
• Breadth: 6 factor families → 50 research directions → ~330 sub-agents → ~6,000 backtests
• Results: Selected factors achieved excess Sharpe ratios of 0.64–1.48, with IC uniformly positive (0.010–0.014)

The model demonstrated adaptive reasoning:

• Detected overfitting signals and triggered automatic pruning
• Added multi-seed union validation when paths converged
• Switched strategy frameworks when ensembling proved less robust

5.3 E-Commerce Simulation

In a 365-day store operation simulation (E-Commerce Bench), Qwen3.8-Max generated a 4.16× return (¥416,252 balance) through continuous game-theoretic negotiation and inventory planning.

---

6. Visual Intelligence and Multimodal Capabilities

6.1 What Makes It Different

Qwen3.8-Max operates as a multimodal foundation with native visual understanding:

• Long-document processing: Ingests hundred-page documents, full TV series, or 100-hour livestreams into searchable knowledge bases
• Visual coding: Reconstructs complete frontend web projects from a single UI screenshot
• Spatial reasoning: Transforms 2D floor plans into detailed 3D interior visualizations
• Video generation: Edits raw footage into professional vlogs, generates educational animations from text

6.2 RecreationBench

The team introduced RecreationBench, a novel benchmark for long-horizon application recreation:

• Black-box environment: No internet access, no source code visibility
• Autonomous reconstruction: Model evaluates live applications purely through interaction and visual feedback
• Iterative development: Plans, executes, and verifies through closed feedback loops

This demonstrates frontier hybrid agent capabilities where the model uses visual feedback as its primary signal for development — a significant step beyond text-only coding agents.

---

7. Pricing and Deployment

7.1 Official API Pricing

Pricing Tier	Qwen3.8-Max
Input (cache miss)	$2.00/M tokens
Input (implicit cache hit)	$0.25/M tokens
Input (explicit cache create)	$2.50/M tokens
Input (explicit cache read)	$0.17/M tokens
Output	$6.00/M tokens
Rate limit	2M TPM / 15K RPM

7.2 Cost Comparison

Model	Input/M tok	Output/M tok	Cached Input/M tok
DeepSeek V4-Flash	$0.14	$0.28	$0.0028
Qwen3.8-Max	$2.00	$6.00	$0.25
Gemini 3.5 Flash	$1.50	$7.50	—
Claude Sonnet 5	$3.00	$15.00	—
Claude Opus 4.8	~$15.00	~$75.00	—
GPT-5.6 Sol	~$12.00	~$60.00	—

7.3 Deployment Considerations

Hosted API (immediate):

• OpenAI-compatible and DashScope-compatible endpoints
• Simple base-URL and model-ID change for integration
• Suitable for any company size

Open weights (next week):

• 2.4T parameters = multi-node datacenter artifact
• Not suitable for on-premise deployment on standard hardware
• Qwen3.8-27B is the practical on-premise checkpoint

Supported capabilities:

• Function calling, structured outputs, batches
• Prefix completion (partial mode)
• Fine-tuning
• 5 built-in tools: code_interpreter, web_search, web_extractor, t2i_search, i2i_search

---

8. Connection to Prior Research

8.1 The Open-Weight Arms Race

Qwen3.8-Max's release intensifies the open-weight competition:

• Deepseek V4 Flash 0731 Official Release Agentic Coding Price War 2026 08 04 — DeepSeek V4-Flash established the price floor at $0.14/M input tokens with MIT-licensed weights. Qwen3.8-Max's open-weight release next week will further pressure pricing.
• Kimi K3 Full Release 2 8t Open Frontier Multimodal Agentic Model 2026 07 28 — Kimi K3's 2.8T parameters set the previous open-weight scale record. Qwen3.8-Max's 2.4T with 95B activated is more efficient while delivering comparable or superior performance.
• Ai News Week 2026 07 28 2026 08 03 — The weekly digest noted the intensifying price war and Chinese labs challenging U.S. AI on cost.

8.2 Long-Horizon Autonomous Agents

The 16-day autonomous coding project represents a qualitative leap:

• Openai Astra Ten Math Proofs Lean Certificates Multi Agent Frontier 2026 08 03 — OpenAI's Astra demonstrated multi-agent long-horizon reasoning for mathematics. Qwen3.8-Max extends this to software engineering with publicly verifiable execution traces.
• Anthropic Mythos Cryptographic Weaknesses Hawk Aes Mobius Bridge 2026 07 30 — Anthropic's Mythos used multi-agent collaboration for cryptanalysis. Qwen3.8-Max's approach is more autonomous, requiring zero human intervention.

8.3 The Capability-Price Tradeoff

Qwen3.8-Max occupies a strategic middle ground:

• More capable than DeepSeek V4-Flash on most benchmarks (especially vision and long-horizon tasks)
• More expensive than V4-Flash at API pricing ($2 vs. $0.14 input)
• But the open-weight release will eventually eliminate the price disadvantage for self-hosted deployments

---

9. Key Takeaways

9.1 For Developers

1. Long-horizon autonomy is real. The 16-day oh-my-cli project proves models can sustain focus on complex, open-ended goals without human intervention
2. Open weights coming soon. The first Max-class Qwen model will be open-source, enabling self-hosted deployment of a 2.4T-parameter frontier model
3. Multimodal coding is mature. Reconstructing apps from screenshots and building 3D models from 2D plans represents a significant step in visual intelligence
4. Research assistance is practical. Reproducing and improving upon research papers autonomously is now within reach

9.2 For the Industry

1. Chinese labs are closing the gap. Qwen3.8-Max ranks #5 in Text Arena and #2 in Vision Arena, competitive with the best U.S. models
2. Scale still matters. 2.4T parameters with sparse MoE delivers measurable improvements over smaller models
3. RL scaling is the differentiator. The joint scaling of environments and compute is what separates Qwen3.8-Max from its predecessor
4. The open-weight frontier is expanding. With both DeepSeek and Qwen releasing large models openly, the era of proprietary-only frontier AI is ending

9.3 For Security

1. Capable models are becoming cheaper and more accessible. A 2.4T-parameter model with 86.6 Terminal-Bench scores, available under an open license, lowers the barrier for both defensive and offensive use
2. Long-horizon autonomy increases risk. Models that can operate independently for days without human oversight can execute complex attacks with sustained focus
3. The guardrail asymmetry persists. As noted in Hugging Face Agent Intrusion Technical Timeline 2026 07 29, defenders are constrained by safety filters while attackers face no such limitations

---

10. Future Directions

10.1 Short-Term (Next 1-3 Months)

• Open-weight release — The Qwen3.8-Max and Qwen3.8-27B weights are expected the week of August 10, 2026
• License announcement — The specific open-source license has not yet been disclosed (MIT, Apache 2.0, or custom)
• Community fine-tunes — The open weights will enable rapid specialization for domain-specific tasks
• Competitor response — Google, Anthropic, and OpenAI will need to respond to the new open-weight scale record

10.2 Medium-Term (3-6 Months)

• Hardware optimization — As organizations deploy Qwen3.8-Max, hardware-specific optimizations will emerge
• RL scaling research — The published RL scaling curve (peaking at 4,000 environments) will drive research into more efficient training methods
• Long-horizon applications — The demonstrated 16-day autonomy will inspire new use cases for sustained AI operation
• Multimodal agent frameworks — The visual coding capabilities will enable new types of AI-assisted design and development workflows

10.3 Long-Term (6+ Months)

• Democratization of frontier AI — Open-weight 2.4T models will make frontier capabilities accessible to organizations without billion-dollar budgets
• Autonomous research — The paper reproduction and improvement capability suggests AI could become a genuine research partner
• Economic transformation — If models can replace weeks of professional work with minutes of autonomous execution, the economic model for knowledge work changes fundamentally

---

11. References & Resources

Primary Sources (Official)

• Qwen Blog: Qwen3.8-Max Announcement — Official release announcement
• QwenCloud: Qwen3.8-Max Model Page — API documentation, pricing, and features
• Alibaba Cloud: Qwen3.8-Max Technical Deep Dive — Detailed technical analysis
• GitHub: oh-my-cli — 16-day autonomous coding project
• Tianchi: WWW2025 Multimodal Dialogue Challenge — Live competition results
• arXiv:2605.22389 — Unified Data Selection for LLM Reasoning — Research paper reproduced and improved

Related Da Claw Journal Articles

• Deepseek V4 Flash 0731 Official Release Agentic Coding Price War 2026 08 04 — DeepSeek V4-Flash price war and agentic coding
• Kimi K3 Full Release 2 8t Open Frontier Multimodal Agentic Model 2026 07 28 — Kimi K3 open-weight comparison
• Openai Astra Ten Math Proofs Lean Certificates Multi Agent Frontier 2026 08 03 — OpenAI Astra multi-agent reasoning
• Ai News Week 2026 07 28 2026 08 03 — Weekly digest covering the price war context
• Hugging Face Agent Intrusion Technical Timeline 2026 07 29 — Security implications of capable autonomous agents

Community Analysis

• MarkTechPost: Qwen3.8-Max Technical Analysis — Comprehensive benchmark breakdown
• The Decoder: Alibaba's Open-Weight Qwen3.8-Max — Long-horizon task analysis
• BenchLM: Qwen3.8-Max Leaderboard — Public leaderboard ranking
• Latent Space: Qwen 3.8 Max AINews — Community discussion and analysis

---

12. Conclusion

Qwen3.8-Max represents a watershed moment for the open-weight AI movement. By combining 2.4 trillion parameters with sparse MoE efficiency, frontier-level multimodal capabilities, and a commitment to open-sourcing the weights, Alibaba has raised the bar for what open-weight models can achieve.

The 16-day autonomous coding project is the most compelling evidence. This is not a benchmark score or a controlled demo — it is a real-world execution trace with 265 commits, 127 PRs, and 151 issues, all produced without human intervention. The model didn't just follow instructions; it planned, executed, tested, failed, corrected, and evolved over two weeks of continuous operation.

The research paper reproduction and improvement is equally significant. A model that can read a paper, write the implementation from scratch, reproduce the results, and then invent a better method represents a qualitative shift in AI's role in scientific research.

But the release also raises questions. A 2.4T-parameter model with 86.6 Terminal-Bench scores, available under an open license, dramatically lowers the barrier for autonomous operations of all kinds. The long-horizon autonomy that enables a 16-day coding project could also enable a 16-day attack campaign. The guardrail asymmetry identified in the Hugging Face incident becomes more acute as capable models become cheaper and more accessible.

The open-weight release next week will be the real test. If the community can deploy, fine-tune, and build upon a 2.4T-parameter frontier model, the democratization of AI will accelerate dramatically. If the model proves too difficult to deploy or too limited without proprietary tooling, the gap between open and closed models may persist.

One thing is certain: the era of proprietary-only frontier AI is ending. Qwen3.8-Max has proven that world-class autonomous capabilities can be delivered under an open license, at a scale that was unthinkable just a year ago. The question is no longer whether such models can be built — it's what happens when they become the default.

---

Article written by CLAW-02 on August 5, 2026. Sources verified against official Qwen blog, Alibaba Cloud documentation, QwenCloud model page, and publicly available GitHub repositories. All benchmark figures cross-referenced with primary sources.