Qwen 3.8 Max: Complete Benchmark Guide vs GPT-5.6, Claude Fable 5, DeepSeek V4, Kimi K3, MiniMax M3, GLM-5.2 & More (August 2026)

By AiCybr Editorial · Published August 5, 2026

What This Article Covers

Qwen 3.8 Max (model ID: qwen3.8-max-preview) is Alibaba’s flagship frontier model, officially unveiled on August 3, 2026 after a Token Plan preview in July. It is a 2.4-trillion-parameter sparse Mixture-of-Experts (MoE) model with approximately 95 billion active parameters per token, a 1-million-token context window, and support for text, images, and video as input.

This guide covers:

• Technical architecture and what Alibaba has — and has not — disclosed
• Head-to-head benchmark comparison against 12+ frontier models (open & closed)
• Token Plan access, pricing, and important usage restrictions
• API pricing comparison across all major providers
• The Qwen model family evolution from 2023 to 2026
• Agentic capabilities and the Qwen-AgentWorld ecosystem
• Open-weights release timeline and hardware requirements
• Practical decision framework: when to use Qwen 3.8 Max vs alternatives
• Sources and further reading

Important: Qwen 3.8 Max was released just two days ago at the time of this writing. While Alibaba has published benchmark results, independent verification is still limited. Vendor-reported numbers are clearly labeled throughout this article. We recommend cross-referencing with live leaderboards like Artificial Analysis, BenchLM.ai, and Arena.AI as independent evaluations grow.

Qwen 3.8 Max at a Glance

Sources: Qwen3.8 launch page, QwenCloud model guide, Reuters launch report

Architecture & Key Innovations

Sparse Mixture-of-Experts (MoE) Design

Qwen 3.8 Max uses a Mixture-of-Experts (MoE) architecture. Instead of running all 2.4 trillion parameters for every token, the model contains many specialized sections called experts. A routing system selects the most relevant experts for each token — Alibaba says approximately 95 billion parameters are active at any given time.

This design can reduce per-token computation compared to a dense 2.4T model, but the full model still has to be stored and made available across the serving system.

What Alibaba Has — and Has Not — Disclosed

Alibaba has confirmed the model’s total size, active parameter count, MoE design, context length, and multimodal capabilities. As of August 5, 2026, it has not yet published a full technical report with:

• Total number of experts
• Experts selected per token
• Exact attention-layer arrangement
• Training-token count and data composition
• Training hardware and total compute
• Quantization formats for downloadable weights

Hybrid Attention (Gated DeltaNet)

Recent Qwen models — including Qwen3-Next — use a hybrid of Gated DeltaNet (linear attention) and standard full attention layers. This hybrid approach:

• Replaces quadratic-complexity self-attention with O(n) linear attention in the majority of layers
• Retains full attention in a subset of layers (approximately 1 in 4) for high-fidelity reasoning
• Interleaves 3 linear layers per 1 full attention layer (3:1 ratio)

Caveat: Alibaba has not published a technical report confirming the exact layer ratio or implementation for Qwen 3.8 Max specifically. The 3:1 ratio and Gated DeltaNet design are established for Qwen3-Next. A Qwen3.8 technical report is needed before these specifics can be stated with precision.

Architecture reference: Qwen3-Next architecture blog

Native Multimodality

Qwen 3.8 Max processes text, images, and video through a unified architecture. This enables document understanding, OCR on complex layouts, video comprehension, and image-grounded reasoning. Note that image generation is a separate capability provided through other Alibaba models (e.g., Wan image models) — it is not a native output of Qwen 3.8 Max itself.

Context Window: 1 Million Tokens

One million tokens is roughly equivalent to several hundred thousand English words. Useful for reviewing large codebases, comparing many documents, processing long agent histories, and analyzing large mixed-media collections.

A large advertised context window does not guarantee perfect recall across the full input. Test the model at the specific lengths you expect to use — 100K, 300K, 600K, near 1M — rather than assuming performance is constant at the limit.

Benchmark Results

Alibaba published a broad benchmark set for Qwen 3.8 Max at launch. All scores below are Alibaba-reported and were not independently reproduced across a common public harness by August 5, 2026.

Alibaba’s Official Launch Benchmark Table

Early Independent Signal: Arena.AI

Reuters reported that Qwen 3.8 Max became the highest-ranked Chinese text model on Arena.AI around its public unveiling, and ranked second globally on the visual leaderboard, behind a Claude Fable 5 variant. Arena results are based on user preferences between anonymous model responses — a useful early signal of real-world response quality, but not a controlled replacement for task-specific benchmarks.

Why Benchmark Numbers Are Not Absolute

Scores change when any of the following changes: model snapshot, prompt template, reasoning effort, temperature, tool access, agent harness, time limits, max output length, number of attempts, or judge model. The harness (software around the model) can materially improve scores even when the underlying model is unchanged. This is why vendor-reported scores should not be placed into a single ranking alongside independently reproduced scores unless evaluation setups are matched.

Full Multi-Model Benchmark Comparison

The tables below use scores from Alibaba’s official launch comparison where available (V†), plus independently sourced third-party data (3P) and each provider’s own official reports (V).

Agentic & Coding Benchmarks

Table key: (V) = vendor-reported by that model’s provider, (V†) = as reported in Alibaba’s Qwen 3.8 Max launch comparison table, (3P) = independent third-party, (3P est.) = third-party estimate. — = no published score. * = Open weights announced, not yet released. TB 2.0 = Terminal-Bench 2.0 (prior benchmark version, not directly comparable to TB 2.1).

SWE-bench Verified vs. Pro: Verified (500 instances) is now largely saturated at the top end (93–97% for leaders). SWE-bench Pro covers more complex enterprise tasks and better differentiates frontier models. Scores are not interchangeable.

Terminal-Bench 2.1: Qwen 3.8 Max (86.6%) is the highest open-weight score. GPT-5.6 Sol leads overall (~88.8% per Alibaba). Note: DeepSeek V4 Pro’s 67.9% was on Terminal-Bench 2.0 — not directly comparable.

OSWorld-Verified: Alibaba claims Qwen 3.8 Max (86.1%) surpasses Claude Fable 5 (~85.0%) and GPT-5.6 Sol (~83.2%). These competitor scores are from Alibaba’s own launch table — treat with caution pending independent verification.

Reasoning & Knowledge Benchmarks

GPQA Diamond: Gemini 3.1 Pro leads at 94.3%, followed by GPT-5.6 Sol (~94.1%). Qwen 3.8 Max and Claude Fable 5 are tied at 92.6% — Qwen is firmly frontier-tier but does not lead on this metric.

HLE (Humanity’s Last Exam): Claude Fable 5 leads at 53.3%, then GPT-5.6 Sol (47.2%), Claude Opus 4.8 (45.7%), and Qwen 3.8 Max at 43.6%. Qwen trails all three proprietary models listed. Note: independently tracked HLE scores for Claude Opus 5 (~83.0%) and Kimi K3 (~80.0%) are from different evaluation versions and not directly comparable.

PaperBench: Qwen 3.8 Max claims 93.0% — the highest listed — above GPT-5.6 Sol (~90.5%) and Claude Fable 5 (~88.8%). All competitor scores are from Alibaba’s launch table.

Instruction-Following Benchmark (IFBench)

Qwen 3.8 Max leads IFBench by a wide margin. All competitor scores are from Alibaba’s launch comparison table — not independently verified.

Multimodal Benchmarks

Competitor scores were not published for these specific tests at research time.

BenchLM Composite Rankings (August 4, 2026)

The #46 composite ranking reflects early evaluation data for a model released 2 days ago. BenchLM scores typically improve as harnesses are tuned. Notably, Qwen 3.8 Max ranks #1 in the Reasoning subcategory — suggesting its strengths may not yet be fully captured by the composite score.

Key Benchmark Observations

Where Qwen 3.8 Max Excels

1. IFBench (82.8%): Leads all models in the comparison — Claude Fable 5 (63.5%) and GPT-5.6 Sol (72.7%) trail significantly on instruction-following precision.

2. PaperBench (93.0%): Highest score on scientific paper analysis per Alibaba’s launch data, above GPT-5.6 Sol (90.5%) and Claude Fable 5 (88.8%).

3. Terminal-Bench 2.1 (86.6%): Highest reported open-weight score. GPT-5.6 Sol leads overall (~88.8%), but Qwen 3.8 Max exceeds Claude Fable 5 (84.6%) and DeepSeek V4 Flash 0731 (82.7%).

4. OSWorld-Verified (86.1%): Claims to surpass both Claude Fable 5 (~85.0%) and GPT-5.6 Sol (~83.2%) on desktop OS interaction — key for computer-use agentic workflows.

5. GPQA Diamond (92.6%): Tied with Claude Fable 5; only 1.5 points behind GPT-5.6 Sol and 1.7 points behind Gemini 3.1 Pro. Firmly frontier-tier for scientific reasoning.

6. Cost: At Token Plan rates (starting $6/month) and with forthcoming pay-as-you-go pricing expected to be competitive, it is dramatically cheaper than Claude Fable 5 ($10/$50 per 1M) and GPT-5.6 Sol ($5/$30 per 1M).

Where Qwen 3.8 Max Falls Short

1. SWE-bench Pro (67.7%): The clearest weakness. Claude Fable 5 scores ~80.0% — a ~12-point gap on complex enterprise coding tasks. GPT-5.6 Sol (~64.6%) and Claude Opus 4.8 (~69.2%) are closer, but Fable 5 sets a significantly higher bar.

2. HLE — Humanity’s Last Exam (43.6%): Trails Claude Fable 5 (53.3%), GPT-5.6 Sol (47.2%), and Claude Opus 4.8 (45.7%) on this extremely hard expert benchmark. Independently tracked frontier models like Claude Opus 5 and Kimi K3 score significantly higher (~80%).

3. BenchLM Composite (#46): Real-world breadth performance is not yet verified to be at the absolute frontier. Expected to improve as evaluation harnesses are tuned.

4. No SWE-bench Verified score: Alibaba has not published a score. Leading models score 93–97% on this widely-tracked coding benchmark.

5. Independent verification pending: Most data is from Alibaba’s own launch table, released 2 days ago. The gap between vendor-reported and independently verified scores typically narrows over weeks.

Access and Pricing

Current Access Route

Alibaba’s QwenCloud documentation lists qwen3.8-max-preview as Token Plan only. It is currently not listed as a standalone pay-as-you-go model in Alibaba Cloud’s public inference pricing table.

Personal Token Plan

Alibaba’s Personal Edition is available in the Singapore region, using rolling Credit limits:

Limited-time promotional prices. Verify the checkout page before subscribing.

Important Usage Restriction

The Personal Token Plan is intended for interactive use inside supported tools (Qwen Code, Claude Code, Cursor, OpenClaw). Personal Plan keys must not be used for production backends, unattended automation, scheduled batch jobs, or non-interactive API workloads. Teams building automated services should use an appropriate commercial or team plan.

API Compatibility

Alibaba provides OpenAI-compatible access and documents Anthropic-compatible endpoints for Token Plan integrations. Compatibility means existing clients can connect by changing the model name, key, and base URL. It does not mean every provider-specific feature behaves identically.

Sources: Personal Token Plan overview, Token Plan quick start, Model Studio pricing

Pricing Compared With Other Frontier Models

Prices verified August 5, 2026. All prices in USD per 1M tokens unless noted. Verify at official provider documentation before production decisions.

Notes: * = Open weights announced, not yet released. † = Gemini 3.1 Pro pricing rises above 200K input tokens ($2→$4 input, $12→$18 output). ‡ = Claude Sonnet 5 introductory pricing through August 31, 2026 (standard: $3/$15). § = GPT-5.6 Sol requests >272K input tokens are billed at 2× input rate and 1.5× output rate.

Official pricing sources: GPT-5.6 Sol, Claude pricing, Gemini 3 pricing, DeepSeek pricing

The Models Compared: Quick Profiles

Closed-Source (Proprietary) Models

GPT-5.6 Sol — OpenAI

• Released: July 9, 2026
• Architecture: Proprietary (tiered: Sol / Terra / Luna)
• Context: 1,050,000 tokens (128K max output; surcharges above 272K input)
• Benchmarks: SWE-bench Verified ~96.2%, Terminal-Bench 2.1 ~88.8%, GPQA ~94.1%
• Pricing: $5 / $30 per 1M tokens (input / output)

Claude Fable 5 — Anthropic

• Released: June 9, 2026
• Architecture: Proprietary
• Context: 1,000,000 tokens (128K max output)
• Benchmarks: SWE-bench Verified ~95.0%, SWE-bench Pro ~80.0%, GPQA ~92.6%, HLE ~53.3%
• Pricing: $10 / $50 per 1M tokens (input / output)

Claude Opus 5 — Anthropic

• Architecture: Proprietary
• Context: 1,000,000 tokens (128K max output)
• Benchmarks: SWE-bench Verified ~96.0%, HLE ~83.0% (3P), BenchLM #1
• Pricing: $5 / $25 per 1M tokens (input / output)

Gemini 3.1 Pro — Google DeepMind

• Released: February 19, 2026 (Preview)
• Architecture: Proprietary
• Context: 2,000,000 tokens (largest in frontier tier)
• Benchmarks: GPQA Diamond 94.3% (industry leader), SWE-bench Verified 80.6%
• Pricing: $2–$4 / $12–$18 per 1M tokens (tiered above 200K input)

Gemini 3.6 Flash — Google DeepMind

• Released: July 21, 2026 (GA)
• Context: 1,000,000 tokens
• Benchmarks: BenchLM 75.3 (leads Google lineup), 17% fewer output tokens vs 3.5 Flash
• Pricing: Budget Flash tier; optimized for speed and cost

Open-Source / Open-Weight Models

DeepSeek V4 Pro — DeepSeek

• Released: April 24, 2026
• Architecture: 1.6T MoE (49B active)
• Context: 1,000,000 tokens | License: MIT
• Benchmarks: SWE-bench Verified 80.6%, SWE-bench Pro ~55.4%, LiveCodeBench 93.5%, GPQA 90.1%
• Pricing: $0.435 / $0.87 per 1M tokens

DeepSeek V4 Flash (0731) — DeepSeek

• Released: April 24, 2026 (updated July 31, 2026)
• Architecture: 284B MoE (13B active)
• Context: 1,000,000 tokens | License: MIT
• Benchmarks: Terminal-Bench 2.1 82.7% (surpassed original V4 Pro on TB 2.1)
• Pricing: $0.14 / $0.28 per 1M tokens

Kimi K3 — Moonshot AI

• Released: July 16, 2026 (weights: July 27, 2026)
• Architecture: 2.8T MoE (104B active, 896 experts); Kimi Delta Attention (KDA) + Attention Residuals
• Context: ~1,048,576 tokens | License: Kimi K3 License (read terms before commercial use)
• Benchmarks: BenchLM #5 (79.8), SWE-bench Verified ~93.4%, HLE ~80.0%
• Pricing: $3 / $15 per 1M tokens

MiniMax M3 — MiniMax

• Released: June 1, 2026
• Architecture: 428B MoE (23B active); MiniMax Sparse Attention (MSA)
• Context: 1,048,576 tokens | License: Open weights
• Benchmarks: GPQA Diamond 88.9%, HLE ~74.0%, SWE-bench Pro ~59.0%
• Pricing: $0.30–$0.60 / $1.20–$2.40 per 1M tokens (tiered)

GLM-5.2 — Zhipu AI (Z.ai)

• Released: June 13, 2026 (weights: June 16, 2026)
• Architecture: ~744–753B MoE (~40B active); IndexShare (2.9× fewer FLOPs at long context)
• Context: 1,000,000 tokens | License: MIT
• Benchmarks: AIME 2026 99.2%, GPQA Diamond 91.2%, SWE-bench Pro 62.1%, Terminal-Bench 2.1 81.0%
• Trained entirely on Huawei Ascend 910B — no NVIDIA dependency
• Pricing: ~$1.40 / $4.40 per 1M tokens

GLM-5 (Base) — Zhipu AI (Z.ai)

• Released: February 11, 2026
• Architecture: 744B MoE (~40B active)
• Context: 200,000 tokens (GLM-5.2 expanded to 1M) | License: MIT
• Benchmarks: SWE-bench Verified 77.8%, AIME 2026 92.7%, GPQA Diamond 86.0%
• Trained on Huawei Ascend chips (first GLM-5 series model; GLM-5.2 improved on all scores)

Llama 4 Maverick — Meta

• Released: April 2026
• Architecture: ~400B MoE (17B active, 128 experts)
• Context: 1,000,000 tokens | License: Llama License
• Note: Behemoth (2T) was never publicly released; Meta pivoted to closed-weight Muse Spark

Grok 4.5 — xAI

• Architecture: Proprietary
• Benchmarks: SWE-bench Verified ~86.6% (3P)

Qwen-AgentWorld: A Separate Research Project

Qwen-AgentWorld is a related but separate Qwen ecosystem project — it is not a built-in component of Qwen 3.8 Max itself.

AgentWorld is a language world model trained to predict how an environment responds to an action (e.g., what a terminal command outputs, how a web page changes after a click), enabling simulated “mental rehearsal” before committing to actions. It covers seven environments:

• MCP tool servers, Search, Terminal, Software engineering
• Web browsers, Desktop operating systems, Android

Alibaba released separate AgentWorld models at 35B-A3B and 397B-A17B scales. The research is an important Qwen direction, but Alibaba has not stated that Qwen 3.8 Max internally runs AgentWorld before every action.

Source: Qwen-AgentWorld

Agentic Coding Capabilities

Autonomous Long-Horizon Coding

Alibaba demonstrated that Qwen 3.8 Max completed a software-engineering project over 16 days in an agentic setting. This shows the model and its surrounding agent system can maintain state, use tools, and continue working across extended tasks.

Long-running agents can accumulate mistakes, misunderstand requirements, modify tests incorrectly, or drift from goals. For serious coding work, pair with: clear requirements, version control, automated tests, security checks, checkpointed progress, and human review at design/release gates.

The Qwen-Agent Ecosystem

The model integrates with the Qwen-Agent framework, providing:

• Tool calling: Structured function calls with multi-turn state management
• Planning: Task decomposition and sequential execution
• Memory management: Long-context retrieval and summarization
• Multi-agent orchestration: Coordinating multiple agent instances

Open Weights: Timeline & Hardware Requirements

Release Status (August 5, 2026)

Until the files and license appear, several questions remain open:

• Which precisions and quantized versions will be published?
• What hardware topology will Alibaba recommend?
• Will the license allow unrestricted commercial deployment?
• Will the open checkpoint match the hosted preview?

Theoretical Weight Memory

These cover model weights only. A real server also needs memory for KV cache, routing buffers, runtime workspace, framework overhead, and concurrent users.

For perspective: 64 H100 GPUs (80 GB each) = 5.12 TB total GPU memory — barely above theoretical FP16 weight size with little room for overhead. Efficient deployment requires lower-precision weights, expert parallelism, and a carefully designed multi-node serving system.

• Kimi K3 (2.8T): ~64× H100/B200 GPUs
• DeepSeek V4 Pro (1.6T): ~32× H100s
• Qwen 3.8 Max (2.4T): estimated ~48–64× H100/B200 for FP16

The Qwen 3.8-27B variant will be the practical choice for most developers — similar to how Qwen3.6-27B has been widely adopted for local deployment.

The Qwen Model Family: A Brief History

When to Use Qwen 3.8 Max

✅ Good Fit

• Agentic workflows requiring terminal and desktop interaction (strong Terminal-Bench and OSWorld scores)
• Instruction-following precision — IFBench leader by a wide margin
• Scientific research tasks requiring strong GPQA-level reasoning
• Multimodal document processing (OmniDocBench 92.1%)
• Chinese + English bilingual production environments
• Teams evaluating future open-weight alternatives to proprietary frontier models
• Token Plan ecosystem users working with Qwen Code, Claude Code, or OpenClaw

⚠️ Consider Alternatives

How to Evaluate Qwen 3.8 Max Properly

• Build a representative test set: 50–200 tasks from real workflows, including difficult cases, incomplete instructions, and failure scenarios
• Use the same environment for every model: Keep prompts, tools, time limits, and success criteria consistent — comparing models across different agent frameworks is not controlled
• Measure beyond answer quality: Track task success rate, correction time, latency, token/Credit consumption, tool calls, retry rate, hallucinations, unsafe actions, and failure recovery
• Test long context directly: Measure retrieval at multiple input sizes (100K, 300K, 600K, near 1M); place key facts at beginning, middle, and end
• Test preview stability: Rerun a fixed benchmark set regularly — Alibaba may update the preview model without a new model ID
• Keep human review gates: For coding, legal, financial, security, or operational work, require human approval before high-impact actions

The Bigger Picture: August 2026 AI Landscape

1. The frontier is crowded. Six or more models now compete for the top tier across different benchmark categories. “Best model” increasingly depends on the specific task.

2. Open weights are catching up. The gap between open-weight and proprietary models averages approximately 4 months in capability. Kimi K3, DeepSeek V4 Pro, and GLM-5.2 all demonstrate near-frontier performance.

3. Pricing compression is accelerating. Open-weight models (DeepSeek from $0.14/1M) put massive pressure on proprietary pricing. Even Qwen 3.8 Max’s Token Plan starting at $6/month undercuts Claude Fable 5 by an order of magnitude for many use cases.

4. Agentic capability is the new differentiator. Terminal-Bench, OSWorld, SWE-bench Pro, and FrontierSWE have replaced MMLU as the benchmarks that matter. Qwen 3.8 Max’s strong Terminal-Bench and OSWorld claims reflect this shift.

5. Benchmark verification matters more than ever. As vendor-reported benchmark inflation becomes common, independent evaluations (Arena.AI user preferences, BenchLM, Artificial Analysis) are increasingly the most trustworthy signal.

Final Assessment

Qwen 3.8 Max is Alibaba’s most ambitious model release to date. Its 2.4-trillion-parameter MoE design, 95B active parameters, 1M context window, and support for text, images, and video place it among the largest frontier models available through a hosted service.

Alibaba’s benchmark results point to particular strength in:

• IFBench (82.8%) — clear leader on instruction following
• PaperBench (93.0%) — highest listed score on scientific paper tasks
• Terminal-Bench 2.1 (86.6%) — highest reported open-weight score
• OSWorld-Verified (86.1%) — strong computer-use agent capability

Its appearance near the top of Arena.AI’s early text and visual rankings provides an encouraging independent signal.

However, important caveats remain:

• SWE-bench Pro (67.7%) lags significantly behind Claude Fable 5 (~80.0%) — a ~12-point gap on complex enterprise coding tasks
• HLE (43.6%) trails Claude Fable 5 (53.3%) and GPT-5.6 Sol (47.2%) on extremely hard reasoning tasks
• No SWE-bench Verified score published by Alibaba
• BenchLM composite (#46) suggests the model may not match the absolute frontier across all tasks
• Open weights not yet released — all claims are based on API testing only
• Most benchmark data is vendor-reported — independent verification is pending
• Preview status — pricing, availability, and model behavior may change

Our recommendation: Qwen 3.8 Max is one of the most interesting new models of August 2026 for experimentation, long-context analysis, and agentic coding. For production deployment, test against real workloads and compare with models that already offer stable APIs or downloadable weights.

The best model is not the one with the largest parameter count or highest vendor chart score. It is the one that completes your actual tasks reliably, at an acceptable cost, within your privacy, deployment, and operational constraints.

Sources and Further Reading

Qwen and Alibaba Cloud

1. Qwen3.8 launch page
2. QwenCloud text-generation model guide
3. QwenCloud introduction and API compatibility
4. Alibaba Cloud Personal Token Plan overview
5. Alibaba Cloud Personal Token Plan quick start
6. Alibaba Cloud Token Plan Harness tools
7. Alibaba Cloud Model Studio inference pricing
8. Qwen3-Next architecture
9. Qwen-AgentWorld

Independent Coverage

10. Reuters: Alibaba unveils Qwen3.8-Max
11. BenchLM.ai leaderboard
12. Artificial Analysis

Alternative Models and Pricing

13. OpenAI GPT-5.6 Sol documentation
14. Anthropic Claude model overview
15. Anthropic API pricing
16. Google Gemini 3 developer guide
17. DeepSeek V4 models and pricing
18. Kimi K3 model repository
19. GLM-5.2 technical launch

Prices, preview terms, and model availability can change. Verify the provider’s current documentation before making a production or purchasing decision. Leaderboard scores are dynamic — recheck before publication.