Qwen3.8-Max

Qwen3.8-Max is built "upon the architectural foundation of Qwen 3.5" and is fully documented through its open-weight edition. The Hugging Face model card and config describe a 92-layer causal language model with hidden size 8192, arranged as 23 repeats of three Gated DeltaNet (linear attention) blocks followed by one Gated Attention block, each feeding a mixture-of-experts layer. Gated DeltaNet uses 128 value heads and 16 query/key heads at head dimension 128; Gated Attention uses 64 query heads and 4 key/value heads at head dimension 256, with rotary embeddings on 64 dimensions. Each MoE layer holds 512 experts and routes every token to 10 of them plus 1 shared expert (expert width 2048). The vocabulary is 248,320 tokens, and multi-token prediction was trained with multiple steps. The checkpoint holds 2,446,182,725,504 BF16 parameters (Alibaba rounds this to 2.4T) with about 95B active per token.

Native context is 262,144 tokens, extensible to 1,010,000; the API serves 1M by default. The API model adds vision input, a non-thinking mode and built-in tools on top of the open language model; Alibaba calls the vision "native" but does not document how the vision path is attached, and the downloadable weights carry no vision component. Training data volume, compute and knowledge cutoff are not published.

Coding (9.0): Alibaba reports Terminal-Bench 2.1 86.6 (behind GPT-5.6 Sol's 88.8, ahead of Claude Opus 4.8's 84.6), SWE-bench Pro 67.7 on its own corrected task set, PaperBench 93.0 and FrontierSWE 73.5, and shows a 16-day autonomous run that built the oh-my-cli project (265 commits, 127 PRs). Independent signal is strong on web development, where LMArena's code board puts it sixth (1671) and the 0902 snapshot fourth (1681), and middling on the text-coding category (1522, level with Claude Sonnet 5).

Reasoning (9.0): GPQA Diamond 92.6 and Humanity's Last Exam 43.6 without tools / 56.2 with tools, all vendor-reported and below Claude Fable 5 and GPT-5.6 Sol on HLE. Artificial Analysis scores the open-weight edition 40 on Intelligence Index v4.3.2, fifth of 114 large open-weight models and 11 points above Qwen3.7-Max.

Math (8.5): no AIME, HMMT or MATH figure is published; the score rests on vendor visual-math and physics results (MathVision 95.2, HiPhO 90.0) and the reasoning results above.

Agentic (9.2): the release's centre of gravity. Vendor figures include OSWorld-Verified 86.1 (above Fable 5's 85.0), Toolathlon Verified 72.5, SkillsBench 70.2, CoWorkBench 74.8 and WideSearch 81.9, plus showcases of a five-day paper reproduction and roughly 330 sub-agents run from one conversation. Alibaba trained it with reinforcement learning across the QwenWork, Claude Code, Codex, OpenClaw and Hermes harnesses, so it behaves consistently inside third-party agent tools.

Long context (9.0): 1M tokens at one flat price, MRCR v2 8-needle at 256K of 92.9 (vendor) against 86.7 for Qwen3.7-Max, and LongBench v2 66.3.

Multilingual (9.0): a Qwen-family strength; bilingual English/Chinese evaluation throughout and the top MTVQA score (56.6) in Alibaba's multilingual text-in-image comparison. Languages count is not stated.

Vision (9.3) and documents/OCR (9.0): the strongest independently confirmed dimension. LMArena's vision board ranks Qwen3.8-Max second of all models (1302), behind only Claude Fable 5 and ahead of Claude Opus 5 and GPT-5.6 Sol. Vendor figures add MMMU-Pro 82.3, ScreenSpot Pro 84.5, VideoMME 90.4, OmniDocBench 1.5 92.1 and CharXiv reasoning 88.4, handling 200-page PDFs and 100-hour video.

Creative writing (8.0): LMArena's overall text board places it 22nd (1481), statistically level with GPT-5.6 Sol and GLM-5.3; Alibaba makes no creative-writing claim.

Instruction following (9.0): IFBench 82.8, the highest in Alibaba's table (GPT-5.6 Sol 72.7).

Function calling (9.0): function calling, structured outputs, parallel tool calls, prefix completion, and OpenAI Chat Completions, OpenAI Responses and Anthropic-compatible endpoints.

Safety calibration (7.5): automatic content moderation on every Qwen Cloud request, but no published safety framework or refusal data.

Real-time data (7.0): built-in web_search, web_extractor and web_fetch tools; web search is not available on the base model in the Frankfurt, Virginia, Tokyo and Hong Kong regions.

Vendor rows are from Alibaba's launch table (2026-08-03) and use Alibaba's own harnesses: Qwen ran Terminal-Bench 2.1 in Claude Code with a 5-hour timeout, while the competitor figures are the best published scores from other harnesses; SWE-bench Pro uses a corrected task set, so it does not match public leaderboards; PaperBench is judged by Claude Opus 4.6. Many other rows in the launch table (QwenSWEBench, CoWorkBench, RecreationBench, QwenVisualOffice and others) are Alibaba in-house benchmarks and are not listed here. Artificial Analysis measures the open-weight edition Qwen3.8-2.4T-A95B (text-only, served on Alibaba's API); its Qwen3.8-Max page was not readable. LMArena figures are as of 2026-09-21.

Artificial Analysis measures the open-weight edition on Alibaba's API at 39.1 output tokens/sec with a 3.02-second time to first token, against a 77.9 tokens/sec median for large open-weight models. That is about one fifth of Qwen3.7-Max's 198.8 tokens/sec and close to Kimi K3 (42.8 tokens/sec). It is also verbose: 170M output tokens to run the Intelligence Index, against a 140M median. Interactive use at the default xhigh effort feels slow; drop reasoning_effort to medium or low, or turn thinking off on the API, for chat and IDE loops, and keep xhigh for background agent runs. Alibaba sells a faster serving mode, qwen3.8-max-prime, only in the China (Beijing) region at $3.301 / $9.902.

At $2 / $6 it matches Grok 4.7, undercuts Claude Sonnet 5 ($2 / $10), Kimi K3 ($3 / $15) and GPT-5.6 Sol ($4 / $20 promotional), and costs more than GLM-5.3 ($1.40 / $4.40) and DeepSeek V4 Pro ($1.32 / $3.96 at peak). Verbosity erodes part of the headline advantage: Artificial Analysis spent $2,633 to run its index on the open-weight edition, against $2,277 for Qwen3.7-Max and $3,658 for Kimi K3.

Faster serving (Prime mode). Alibaba also sells qwen3.8-max-prime (listed 2026-09-23): the same model with the same capabilities and limits, served at 1.5–2× the standard output speed, at twice the standard price ($3.301 input / $9.902 output per 1M tokens in China (Beijing); $4 / $12 on OpenRouter). Prime has no batch discount or free quota. Source: Alibaba Cloud Model Studio pricing and Prime mode pages; OpenRouter qwen/qwen3.8-max-prime.

Managed API. Qwen Cloud and Alibaba Cloud Model Studio serve qwen3.8-max and the pinned snapshot qwen3.8-max-0902 in six regions: China (Beijing), Singapore, Germany (Frankfurt), US (Virginia), Japan (Tokyo) and Hong Kong. The region fixes where request data is stored; the deployment scope fixes where inference runs. Singapore uses International scope, and Frankfurt, Virginia, Tokyo and Hong Kong use Global scope, which schedules inference in a worldwide pool. Workspace-dedicated endpoints carry a 99.9% SLA and a 3,600-second timeout; the older DashScope domains stop receiving new features after 2026-09-30. OpenRouter routes qwen3.8-max to Alibaba only. No Western hyperscaler lists it.

Open weights. Qwen3.8-2.4T-A95B is on Hugging Face under the Qwen3.8-Max License. It is text-only and always thinks; vision, non-thinking mode, the 1M default context and built-in tools stay API-only. Weight footprints: 4.9 TB in BF16, 2.5 TB in Qwen's official FP8 release, and 1.44 TB in NVIDIA's NVFP4 build, with AMD MXFP4 and Unsloth GGUF variants also published. That is multi-node or multi-GPU-server territory, not a single box. Qwen supplies vLLM, SGLang and TokenSpeed recipes and recommends temperature 1.0, top_p 0.95, top_k 20. Alibaba also serves the open edition as qwen3.8-2.4t-a95b at $2 / $6, and Together, DeepInfra, Novita, SiliconFlow, Modal and Venice host it at the same price.

License, exactly. The Qwen3.8-Max License grants MIT-style rights to use, copy, modify, distribute, sell, host and fine-tune, with two conditions. First, if the model or a derivative powers a commercial product or service with more than 100,000,000 monthly active users or more than US$20,000,000 in monthly revenue, the model name must be prominently displayed in that product's user interface. Second, if the licensee or any affiliate runs a Model-as-a-Service or AI Work Assistant business and the licensee's and affiliates' aggregate revenue exceeds US$50,000,000 in any consecutive 12 months, it must obtain a separate license from Qwen before any commercial use. Model-as-a-Service means giving third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data; merely relaying requests to models hosted by others does not count. An AI Work Assistant is an independent product primarily for AI-assisted coding or office productivity (Qwen's examples: Qoder, QwenWork); single-purpose tools, assistants for other domains, and assistants that are a feature of a product with another primary purpose are excluded. The second condition does not apply to internal use that makes neither the model, its outputs nor its capabilities available to any third party. The revenue test counts total company revenue, not revenue from the AI business. License questions go to model-business@notice.qwencloud.com.

Qwen Cloud screens every API request's input and output with automatic content moderation, encrypts traffic with TLS 1.2+ and stored account data with AES-256, and offers audit and access logs. Alibaba publishes no safety framework, no refusal or hallucination figures, and no SOC 2, ISO 27001, HIPAA or GDPR attestation on the Qwen Cloud or Model Studio model pages. Its data-security page does not address whether API inputs are used for training, so confirm that in the service terms before sending regulated data. Global-scope regions store data in-region but may run inference elsewhere; teams with strict residency needs should use a region whose scope has fixed geographic boundaries, or self-host the open weights, which removes Alibaba from the data path entirely.

Qwen Cloud and Model Studio expose OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols, so the OpenAI and Anthropic SDKs in Python and TypeScript work unchanged, alongside Alibaba's DashScope SDK. Alibaba documents setup for Claude Code, Codex, Qoder CLI, Qwen Code and OpenClaw, and its Token Plan subscription adds Cline, Cursor, OpenCode and Kilo CLI. Qwen-MM-Plugins, released with the model, adds image, video, memory and CAD/Blender tooling to existing agent frameworks. The open weights run on vLLM, SGLang and TokenSpeed, and NVIDIA, AMD, Unsloth and Red Hat have published quantized builds. Adoption is growing: about 55,000 monthly Hugging Face downloads for a 2.4T-parameter checkpoint, six third-party API hosts, and more than 16,000 LMArena text votes.