Qwen3.8-Max

What it is. Qwen3.8-Max is Alibaba’s flagship API model: a 2.4T-parameter MoE (95B activated) with a 1M-token context window, released August 2026 (0902 snapshot dated 2026-09-02). Why it matters. It is the closed-weight flagship of the Qwen line, positioned against DeepSeek-V4-Pro and Kimi K3, and one of the strongest coding models in the database on vendor numbers. Key characteristics. 1M context (991,808 max input), 131,072 max output, 262,144 max chain-of-thought, image/text/video input with text output, thinking and non-thinking modes, function calling, structured outputs, web search, prefix completion, context caching and batch inference. What a professional should know. It is a closed API model — the open Qwen3.8-2.4T-A95B weights are a different, text-only, thinking-only product — and fine-tuning is not supported. Prices differ by region.

Architecture and parameters

Qwen3.8-Max is a 2.4T-parameter MoE with 95B activated (~4.0%), 512 experts (10 routed + 1 shared per token), 92 layers, and a Gated DeltaNet + Gated Attention hybrid. The DeltaNet/linear-attention component is Alibaba’s long-context answer, analogous to Kimi’s KDA and MiniMax’s MSA — it replaces part of the quadratic attention stack so 1M context stays economical. See Mixture-of-Experts and the MoE architectures research.

What the context window actually means

The 1M context (991,808 max input) with a 131,072 output cap places Qwen3.8-Max firmly in the long-input-analysis camp rather than long-form generation. The 262,144 max chain-of-thought is a notable documented ceiling — long internal reasoning can run nearly twice the output cap. See the context-window research.

Pricing implications

Singapore $2 input / $6 output per 1M tokens; Beijing and other Global regions $1.65 / $4.951. The 3x input-to-output ratio and ~18% region spread are the operative facts: where you are billed from changes cost meaningfully, and output tokens dominate. Context caching and batch inference (Beijing only) are the levers for reducing effective cost. See the DeepSeek comparison and GLM-5.3 comparison.

API, coding, and agent implications

The 0902 snapshot upgraded coding and vision. Vendor-reported Terminal-Bench 2.1 (86.6), SWE-bench Pro (67.7), GPQA Diamond (92.6), HLE 43.6 (56.2 with tools) and MRCR v2 256K (92.9) describe a strong all-round flagship with particular strength in long-context retrieval (MRCR). Function calling, structured outputs and web search make it a complete agent substrate. No fine-tuning is a hard constraint for domain-specialization strategies. See Qwen agent and Qwen Code.

Open weights and license

Closed weight, proprietary license, no self-hosting. The open Qwen3.8-2.4T-A95B exists but is explicitly not the same product: text-only and thinking-only. Teams that need open weights for the Qwen line must accept those differences or use the API. See open weight vs API.

Benchmark interpretation

All five rows are vendor-reported from the Qwen3.8-Max column of the model card. GPQA Diamond 92.6 and HLE 43.6 track the flagship reasoning ceiling; MRCR v2 92.9 is a 256K long-context retrieval metric, not a full-1M measure; Terminal-Bench 86.6 is an agentic-coding signal. Scores are from the vendor card, not independent re-measurement.

What the benchmarks do not prove

Vendor-card numbers do not independently verify Qwen3.8-Max against competitors, and cross-vendor comparison is invalid without matching benchmark versions and harnesses. The exact API release date is also unstated — only the 0902 snapshot date is known — so version-drift caveats apply. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: general flagship work — coding, reasoning, long-context retrieval, multimodal input analysis, and agent workflows needing function calling and structured output. Less suited: fine-tuning or custom-domain specialization (unsupported), open-weight/self-hosting requirements, and budget work better served by the flash tier.

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	High
Structured API workflows	High
Agent orchestration	High
Local self-hosted deployment	No evidence
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	High

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates Qwen3.8-Max is Alibaba’s most complete closed flagship: competitive coding scores, a full agent toolkit, six serving regions, and a 1M window — at the cost of closed weights, no fine-tuning, and a region-dependent price. Its long-context strength (MRCR 92.9) and Gated DeltaNet architecture position it for retrieval-heavy and long-document workloads, while the open A95B’s text-only/thinking-only limits mean Alibaba keeps the truly multimodal flagship behind the API. For API-only buyers it is a strong generalist; for fine-tuning or self-hosting teams it is the wrong tool.

Market position and outlook

China AI Hub analysis indicates Qwen3.8-Max is the centerpiece of Alibaba’s hybrid strategy: the database records Alibaba as keeping its two API flagships (Qwen3.8-Max and Qwen3.8-Flash) closed while releasing one open-weight model (Qwen3.8-2.4T-A95B), a pattern distinct from the fully-open Zhipu/DeepSeek and the fully-closed ByteDance. The six documented serving regions — Beijing, Singapore, Hong Kong, Frankfurt, US-Virginia and Tokyo — are the broadest regional footprint in the collection, which matters for data-residency and latency-sensitive international deployments. The ~18% region price spread ($1.65 vs $2.00 input) means region selection is itself a cost decision, and the Gated DeltaNet hybrid positions the model for the retrieval-heavy workloads where its MRCR v2 score is strongest. See the state of China’s AI models and MoE architectures research.

What is uncertain

• The exact API release date is not stated; only the 0902 snapshot date (2026-09-02) is known.
• Benchmark scores come from the vendor model card and are not independently verified.
• The open Qwen3.8-2.4T-A95B weights are a different product (text-only, thinking-only), not this model.

Sources

• Model Studio — qwen3.8-max model detail
• Model Studio — model pricing
• Hugging Face model card — Qwen3.8-2.4T-A95B

Labels used above: Official fact (from primary sources), Vendor-reported claim (benchmark scores published by Alibaba), and China AI Hub analysis (our synthesis, always introduced as such). No third-party benchmark evidence is currently recorded for this model.