Qwen3.8-Max: The First Open-Source Flagship, Hybrid DeltaNet MoE Architecture, and the 10-Day Autonomous Coding Benchmark
On August 26, 2026, Alibaba's Qwen team released Qwen3.8-Max — the first open-sourced Max-class frontier model with 2.4T parameters (95B activated), achieving 86.6% on Terminal Bench 2.1, 56.6% on DeepSWE 1.1, and 93.0% on PaperBench, while introducing the Hybrid Gated DeltaNet + Gated Attention architecture at scale and a preview of the Qwen4 architecture via Qwen3.8-Flash-Next.
Qwen3.8-Max: The First Open-Source Flagship, Hybrid DeltaNet MoE Architecture, and the 10-Day Autonomous Coding Benchmark
Executive Summary
On August 26, 2026, Alibaba's Qwen team released Qwen3.8-Max, marking the first time a Max-class frontier model's weights have been open-sourced. With 2.4 trillion total parameters and 95 billion activated per token (MoE), Qwen3.8-Max represents a fundamental shift in the open-weight landscape: for the first time, the community can self-host a model that competes with the best closed-source frontier models on agentic coding, professional knowledge work, and long-horizon autonomous tasks.
The release delivers across three dimensions: architectural maturity, benchmark dominance, and practical deployment. Architecturally, Qwen3.8-Max builds on the hybrid Gated DeltaNet + Gated Attention design introduced in Qwen3.5, now scaled to 92 layers with 512 experts (10 routed + 1 shared), 248K vocabulary, and native 256K context extensible to 1M tokens. On benchmarks, it achieves 86.6% on Terminal Bench 2.1, 56.6% on DeepSWE 1.1, 93.0% on PaperBench, and 82.8% on IFBench — competing directly with Claude Fable 5 and GPT-5.6 Sol on coding tasks while leading on scientific research and instruction following.
The release package includes three models: Qwen3.8-2.4T-A95B (the flagship MoE, FP8-quantized), Qwen3.8-27B (a dense multimodal model with vision and video understanding), and Qwen3.8-Flash-Next (an experimental preview of the Qwen4 architecture featuring Qwen Sparse Attention, Gated Residual, and N-gram Embedding). API pricing is set at $2/M input and $6/M output — roughly 5× cheaper on input and 8.3× cheaper on output than the US frontier models (GPT-6 Astra at $10/$50, Claude Fable 5.1 at $10/$50).
This article completes the September 2026 frontier model survey alongside Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17, Gpt 6 Astra Critical Cyber Capabilities Computer Use Frontier Alignment Monitorability 2026 09 16, and Gemini 3 8 Flash Agentic Coding Cyber Security Fairwind Program 2026 09 15.
1. The Open-Source Flagship: A Historic Release
1.1 First Max-Class Open Weights
The Qwen3.8 release represents a watershed moment for the open-weight community. Previous open-source models have been either smaller-scale (Qwen3.7-Plus at ~397B) or efficiency-focused (DeepSeek V4-Flash at 284B). Qwen3.8-Max is the first time a model with trillion-parameter scale and frontier-level capabilities has been released with open weights.
| Model | Total Params | Activated | Open Weights? | Release Date |
|---|---|---|---|---|
| Qwen3.8-Max | 2.4T | 95B | Yes (FP8) | Aug 26, 2026 |
| Qwen3.7-Plus | ~397B | ~17B | Yes | July 2026 |
| DeepSeek V4-Flash | 284B | 13B | Yes | May 2026 |
| DeepSeek V4-Pro | 1.6T | 49B | Yes | May 2026 |
| Claude Fable 5.1 | Undisclosed | Undisclosed | No | Sep 1, 2026 |
| GPT-6 Astra | Undisclosed | Undisclosed | No | Sep 3, 2026 |
| Gemini 3.8 Flash | Undisclosed | Undisclosed | No | Sep 2, 2026 |
The FP8 quantization (block size 128) delivers performance "nearly identical to the original model" according to the Qwen team, making it practical for deployment on consumer and enterprise hardware.
1.2 The Qwen3.8 Family
The release comprises three distinct models, each serving a different deployment scenario:
| Model | Type | Params | Activated | Multimodal | Context |
|---|---|---|---|---|---|
| Qwen3.8-2.4T-A95B | MoE (Flagship) | 2.4T | 95B | Text only | 256K → 1M |
| Qwen3.8-27B | Dense | 27B | 27B | Text + Image + Video | 256K → 1M |
| Qwen3.8-Flash-Next | MoE (Experimental) | 125B + 51B n-gram | 6B | Text + Image + Video | 256K → 1M |
2. Architecture: Hybrid Gated DeltaNet + Gated Attention at Scale
2.1 The Hybrid Design
Qwen3.8-Max continues the architectural evolution started with Qwen3.5: a hybrid of Gated DeltaNet (linear attention) and Gated Attention (sparse KV attention) within a Mixture-of-Experts framework. This design combines the O(n) scalability of linear attention with the precision of sparse attention for critical tokens.
Key architectural specifications:
| Component | Specification |
|---|---|
| Hidden Dimension | 8,192 |
| Layers | 92 |
| Hidden Layout | 23 × (3 × Gated DeltaNet → MoE) → (1 × Gated Attention → MoE) |
| Gated DeltaNet Heads | 128 for V, 16 for QK |
| Gated Attention Heads | 64 for Q, 4 for KV (GQA) |
| Head Dimension | 128 (DeltaNet), 256 (Attention) |
| RoPE Dimension | 64 |
| Experts | 512 total, 10 routed + 1 shared |
| Expert Intermediate Dim | 2,048 |
| Vocabulary | 248,320 (padded) |
| MTP | Multi-token prediction, multiple steps |
The 3:1 ratio of DeltaNet to Attention layers (69 DeltaNet layers vs. 23 Attention layers) means the model primarily relies on linear attention for long-range dependencies, falling back to sparse attention for tasks requiring precise token-level reasoning.
2.2 Comparison with Prior Architectures
| Feature | Qwen3.5 | Qwen3.7 | Qwen3.8-Max | DeepSeek V4 |
|---|---|---|---|---|
| Architecture | Hybrid DeltaNet+Attention | Hybrid DeltaNet+Attention | Hybrid DeltaNet+Attention | DSA + Compressed Attention |
| Layers | 60-80 range | 60-80 range | 92 | 60 |
| Experts | 128-256 | 256 | 512 | 256 |
| Routed Experts | 8-10 | 8-10 | 10 + 1 shared | 8 |
| Max Context | 128K-256K | 256K | 256K → 1M | 1M |
| MTP | Yes | Yes | Multi-step | Yes |
2.3 The Qwen4 Preview: Qwen3.8-Flash-Next
Alongside the flagship, Qwen released Qwen3.8-Flash-Next — an experimental model previewing the architecture for Qwen4. This introduces four major innovations:
- Qwen Sparse Attention (QSA): Replaces Gated Attention with micro-block-level selection rather than token-level, significantly reducing long-context latency.
- Gated Residual: Element-wise data-dependent read gates and per-branch scalar write gates modulate information flow through widened residual streams.
- N-gram Embedding: 51 billion parameters indexed by bigrams/trigrams at layer 2, providing a parameter scaling axis that is memory-efficient and amenable to offloading.
- Tailored Training Recipe: Muon and AdamW optimizers applied to specific weight categories, with no batch-size warmup (starting directly at target batch size).
The Flash-Next model achieves 58.7% on DeepSWE 1.1 and 62.5% on SWE-bench Pro with only 6B activated parameters (plus 51B n-gram lookup) — rivaling models with 5-10× more activated parameters.
3. Benchmark Performance: Competing with the US Frontier
3.1 Coding & Agentic Benchmarks
Qwen3.8-Max's most impressive results are in agentic coding, where it competes directly with Claude Fable 5 and GPT-5.6 Sol:
| Benchmark | Qwen3.8-Max | Qwen3.7-Max | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 86.6% | 74.5% | 84.6% | 88.8% | 84.6% |
| SWE-bench Pro | 67.7% | 60.6% | 80.0% | 64.6% | 69.2% |
| DeepSWE 1.1 | 56.6% | 21.6% | 70.0% | 73.0% | 59.0% |
| FrontierSWE | 73.5% | 40.7% | 88.8% | — | 70.0% |
| NL2Repo-Bench | 55.9% | 47.2% | 69.4% | — | 69.4% |
| PaperBench | 93.0% | 64.8% | 88.8% | 90.5% | 80.3% |
| QwenSWEBench | 80.7% | 63.4% | 86.3% | 73.5% | 84.0% |
The PaperBench result of 93.0% is particularly notable — it leads all compared models including GPT-5.6 Sol (90.5%) and Claude Fable 5 (88.8%), suggesting Qwen3.8-Max has superior capability in scientific paper comprehension and code-generation-from-research tasks.
3.2 Professional Knowledge Work
On professional domain benchmarks, Qwen3.8-Max shows strong performance:
| Benchmark | Qwen3.8-Max | Qwen3.7-Max | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| CoWorkBench | 74.8% | 64.6% | 75.9% | 71.5% |
| JobBench | 53.4% | 31.3% | 57.4% | 45.4% |
| SkillsBench | 70.2% | 61.2% | 70.9% | 73.5% |
| PLawBench | 73.2% | 58.9% | 70.2% | 72.3% |
| PRBench-Finance | 58.3% | 46.8% | 55.8% | 55.5% |
| HealthBench | 60.2% | 54.5% | — | 55.3% |
On PLawBench (legal reasoning) and PRBench-Finance, Qwen3.8-Max leads all compared models, suggesting particular strength in Chinese-language legal and financial domains — though English performance remains competitive.
3.3 General Capabilities
| Benchmark | Qwen3.8-Max | Qwen3.7-Max | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| IFBench | 82.8% | 79.1% | 63.5% | 72.7% |
| GPQA Diamond | 92.6% | 92.4% | 92.6% | 94.1% |
| HLE | 43.6% | 41.4% | 53.3% | 47.2% |
| HLE w/ tools | 56.2% | 53.5% | 64.5% | 58.0% |
| MRCR v2 256K | 92.9% | 86.7% | — | 93.8% |
The IFBench score of 82.8% is a standout — it leads all compared models by a significant margin, including GPT-5.6 Sol (72.7%) and Claude Fable 5 (63.5%). This suggests Qwen3.8-Max has exceptional instruction-following capabilities.
4. The 10-Day Autonomous Coding Demo
4.1 What Was Achieved
The Qwen team demonstrated Qwen3.8-Max's capabilities through a 10+ day autonomous coding run where the model:
- Started with minimal inputs: a basic task description, a stub RTL (Register Transfer Level) workspace with empty module templates, and an evaluation script
- Operated completely autonomously, performing requirement collection, issue dispatch, code generation, verification, and self-repair
- Built a self-evolving harness that continuously improved its own evaluation pipeline
- Delivered a complete hardware design project spanning the full development lifecycle
This demonstrates a level of long-horizon autonomous agency that was previously only claimed by closed-source models like Claude Fable 5.1 and GPT-6 Astra.
4.2 Significance
The RTL hardware design domain is particularly challenging because it requires:
- Deep domain knowledge of digital circuit design
- Precise syntax (Verilog/SystemVerilog)
- Iterative verification and synthesis
- Understanding of timing constraints and resource utilization
Achieving this autonomously over 10+ days — with self-repair and continuous improvement — represents a significant milestone for open-source models.
5. Deployment & Implementation
5.1 Self-Hosted Deployment
Qwen3.8-2.4T-A95B is available in FP8 format on Hugging Face, compatible with major inference frameworks:
# vLLM deployment
pip install vllm
vllm serve Qwen/Qwen3.8-2.4T-A95B-FP8 \
--tensor-parallel-size 8 \
--max-model-len 1048576 \
--quantization fp8
# SGLang deployment
pip install sglang
python -m sglang.launch_server \
--model Qwen/Qwen3.8-2.4T-A95B-FP8 \
--port 30000 \
--mem-fraction-static 0.9
Hardware requirements (estimated for FP8):
- Minimum: 8× NVIDIA H100 (80GB) for 256K context
- Recommended: 8× NVIDIA H100 (80GB) for 1M context with caching
- TokenSpeed and KTransformers offer optimized kernels for the hybrid architecture
5.2 API Deployment
Qwen Cloud provides managed inference with the following specifications:
| Parameter | Value |
|---|---|
| Max Input | 991K tokens |
| Max Output | 131K tokens |
| Max Reasoning | 262K tokens |
| Context Window | 1M tokens |
| TPM | 2M tokens/minute |
| RPM | 15K requests/minute |
5.3 Reasoning Effort Control
Qwen3.8-Max supports three reasoning effort levels:
from openai import OpenAI
client = OpenAI(
base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
api_key="your-api-key"
)
# High-effort reasoning for complex tasks
response = client.chat.completions.create(
model="qwen3.8-max",
messages=[{"role": "user", "content": "Design a distributed database sharding strategy"}],
extra_body={
"reasoning_effort": "xhigh", # xhigh (default), medium, low
"chat_template_kwargs": {
"enable_thinking": True,
"preserve_thinking": True,
}
}
)
| Effort Level | Use Case | Estimated Latency | Cost Impact |
|---|---|---|---|
| xhigh | Complex reasoning, research, coding | Highest | ~3× base cost |
| medium | General tasks, balanced accuracy/speed | Moderate | ~1.5× base cost |
| low | Simple queries, rapid iteration | Lowest | Base cost |
6. Pricing Analysis: The Open-Source Advantage
6.1 API Pricing Comparison
| Model | Input ($/M) | Output ($/M) | Cache ($/M) | Context |
|---|---|---|---|---|
| Qwen3.8-Max | $2.00 | $6.00 | $0.25 | 1M |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 | 200K |
| GPT-6 Astra | $10.00 | $50.00 | $2.50 | 200K |
| Gemini 3.8 Flash | $0.75 | $3.75 | — | 1M |
| DeepSeek V4-Pro | $1.50 | $7.50 | — | 1M |
Qwen3.8-Max is 5× cheaper on input and 8.3× cheaper on output than the US frontier models (Fable 5.1, GPT-6 Astra), while offering a 1M context window (vs. 200K for those models). The cache price of $0.25/M matches Anthropic's rate, making it highly cost-effective for long-context workloads with shared prefixes.
6.2 Self-Hosted Cost Estimate
For organizations with existing GPU infrastructure, self-hosting Qwen3.8-Max can be significantly cheaper than API calls:
- 8× H100 (80GB): ~$3-4/hour on major cloud providers
- Estimated throughput: 50-100 tokens/sec (FP8, 8-way TP)
- Cost per 1M output tokens: ~$1-2 (vs. $6 API price)
- Break-even: ~10M output tokens/month
For high-volume agentic workloads (where output tokens dominate), self-hosting becomes economically attractive at relatively low volume thresholds.
7. Integration with Prior Research
7.1 Positioning in the September 2026 Landscape
Qwen3.8-Max completes the picture of the September 2026 frontier model release wave:
| Dimension | Claude Fable 5.1 | GPT-6 Astra | Gemini 3.8 Flash | Qwen3.8-Max |
|---|---|---|---|---|
| Release Date | Sep 1 | Sep 3 | Sep 2 | Aug 26 |
| Open Weights | No | No | No | Yes |
| Agentic Coding | 77.9% OSWorld | 72.6% OSWorld | 73.7% DeepSWE | 86.6% Terminal Bench |
| Scientific | Protein design, Venus mapping | Math (98% FrontierMath T4) | Cybersecurity | PaperBench 93% |
| Cybersecurity | Mythos 5.1 (restricted) | Critical threshold (Daybreak) | Flash Cyber (Fairwind) | Not emphasized |
| Pricing | $10/$50 | $10/$50 | $0.75/$3.75 | $2/$6 |
| Context | 200K | 200K | 1M | 1M |
See Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 for the Anthropic release, Gpt 6 Astra Critical Cyber Capabilities Computer Use Frontier Alignment Monitorability 2026 09 16 for OpenAI, and Gemini 3 8 Flash Agentic Coding Cyber Security Fairwind Program 2026 09 15 for Google.
7.2 The Open-Source vs. Closed-Source Divide
Qwen3.8-Max narrows but does not eliminate the gap between open and closed models:
- Where Qwen leads: Instruction following (IFBench), scientific paper comprehension (PaperBench), legal/financial reasoning (PLawBench, PRBench-Finance), price-to-performance ratio
- Where closed models lead: Cybersecurity capabilities (all three US models have specialized cyber variants), computer use (OSWorld), biology/protein design (Mythos 5.1), monitorability/safety research
The Qwen3.8-Flash-Next architecture preview suggests the gap may narrow further with Qwen4, particularly in long-context efficiency (QSA) and parameter scaling (n-gram embedding).
8. Key Takeaways
-
Open-source frontier is real: Qwen3.8-Max proves that trillion-parameter MoE models can be open-sourced with competitive performance, democratizing access to frontier capabilities.
-
Hybrid architecture matures: The Gated DeltaNet + Gated Attention design, now at 92 layers with 512 experts, represents the most sophisticated open-source architecture to date.
-
Price disruption: At $2/$6 per million tokens, Qwen3.8-Max makes frontier-level agentic coding accessible to organizations that previously could only afford mid-tier models.
-
Long-context advantage: The 1M context window (vs. 200K for Fable 5.1 and GPT-6 Astra) is a practical advantage for document-heavy workflows.
-
Qwen4 preview is promising: The Flash-Next architecture (QSA, Gated Residual, N-gram Embedding) suggests significant efficiency gains coming, potentially enabling frontier performance with far fewer activated parameters.
-
Cybersecurity gap remains: Unlike the US models, Qwen3.8-Max does not emphasize cybersecurity capabilities or offer a restricted cyber variant, representing a potential area for future development.
9. References & Resources
Official Sources
- Qwen3.8-Max Blog Post — Official announcement and capability overview
- Qwen3.8-2.4T-A95B-FP8 on Hugging Face — Model card with full benchmark tables
- Qwen3.8-27B on Hugging Face — Dense multimodal model card
- Qwen3.8-Flash-Next on Hugging Face — Qwen4 architecture preview
- Qwen Cloud Model Page — API specifications and features
- Qwen Cloud Pricing — API pricing details
Inference Frameworks
Related Research
- Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 — Anthropic's September release
- Gpt 6 Astra Critical Cyber Capabilities Computer Use Frontier Alignment Monitorability 2026 09 16 — OpenAI's GPT-6 generation
- Gemini 3 8 Flash Agentic Coding Cyber Security Fairwind Program 2026 09 15 — Google's iterative Flash strategy
10. Future Directions
10.1 Qwen4 Architecture
The Qwen3.8-Flash-Next preview signals several directions for Qwen4:
- Qwen Sparse Attention (QSA): Micro-block-level selection could dramatically reduce latency for long-context agentic workloads
- N-gram Embedding: 51B parameters for pattern matching without compute cost could enable massive scaling on memory-constrained hardware
- Gated Residual: Finer-grained information flow control may enable deeper models without training instability
10.2 Community Adoption
With open weights available, expect rapid community adoption:
- Quantized variants: GGUF, AWQ, and GPTQ versions for consumer hardware
- Fine-tunes: Domain-specific adaptations for legal, medical, and financial use cases
- Benchmarking: Independent evaluations beyond Qwen's own benchmark suite
- Integration: Support in LangChain, LlamaIndex, and agentic frameworks
10.3 Competitive Response
The release of a 2.4T open-weight model may pressure other labs:
- Meta: Potential acceleration of Llama 4 development with MoE architecture
- DeepSeek: Possible V4.5 or V5 release responding to Qwen's capability leap
- US labs: May reconsider open-weight strategies for mid-tier models
10.4 Watch For
- Qwen4 official release (expected Q1 2027 based on the Flash-Next preview timeline)
- Cybersecurity variant of Qwen3.8-Max (if Qwen follows the US lab pattern)
- Multi-modal flagship combining the 2.4T MoE with vision capabilities
- Community benchmarks validating the 10-day autonomous coding claims
Citation:
@misc{qwen38,
title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
url = {https://qwen.ai/blog?id=qwen3.8},
author = {{Qwen Team}},
month = {August},
year = {2026}
}
🔗 Referenced by
- 🔬Grok 4.7: Long-Running Agents, Frontier Coding at $2/M, and xAI's New Safeguard Stack2026-09-25T00:00:00.000Z
- 🔬GPT-6 Sol and Luna: 50% Price Cuts, Agentic Coding at Scale, and the New Cost-Efficiency Frontier2026-09-24T00:00:00.000Z
- 🔬Claude Opus 5.5: Fable 5.1 Performance at 40% Less Cost, the First Release After the Pacing Call2026-09-23T00:00:00.000Z
- 🔬StepFun Step 5 Preview: 600B Sparse MoE, 27B Active Parameters, and the Pareto Frontier for Agentic Coding2026-09-21T00:00:00.000Z
- 📅September 18: Qwen3.8-Max — First Open-Source Flagship, Hybrid DeltaNet MoE, and the 10-Day Autonomous Coding Benchmark2026-09-18T00:00:00.000Z
- 📚Wiki Index2026-06-17T00:00:00.000Z
- 📚Wiki Log2026-06-17T00:00:00.000Z
- 📚Frontier Models & Benchmarks
- 📚Qwen