Loading page
Alibaba
Qwen3.5 397B A17B
Qwen3.5 397B A17B is a large mixture-of-experts language model from Alibaba's Qwen series, with roughly 397B total parameters and about 17B active per token.
Input from
$0.450 / 1M tokens
across 8 providers
API Pricing
Cheapest on Deep Infra — 24% below avg| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.450 | $3.00 | $0.220 | |
| $0.450 | $3.00 | - | |
| $0.550 | $3.50 | $0.225 | |
| $0.600 | $3.60 | - | |
| $0.600 | $3.60 | - | |
| $0.600 | $3.60 | - | |
| $0.688 | $4.13 | - | |
| $0.830 | $5.55 | - |
Prices updated daily. Last check: Sep 22, 2026
Qwen3.5 397B A17B pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
21.4 / 100
Output Speed
83.3 t/s
Latency (TTFT)
1.7s
Reasoning & Knowledge
- GPQA Diamond86.1%
- Humanity's Last Exam19.8%
Agentic & Tool Use
- Terminal-Bench Hard35.6%
- τ²-bench83.9%
Instruction & Long Context
- IFBench51.6%
- Long-Context Reasoning64.3%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Alibaba
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Mixture-of-experts design activates roughly 17B parameters per token out of ~397B total, reducing per-token compute relative to a dense model of comparable size
- Measured output throughput of about 95 tokens per second in Artificial Analysis testing, which is competitive for a model of this parameter class
- Part of Alibaba's Qwen series, which is widely served across multiple third-party inference providers, giving buyers alternatives on price and region
- Large total parameter count provides substantial capacity for knowledge-heavy and reasoning-heavy prompts
- Qwen models have historically had strong Chinese-English bilingual coverage, making the family a common choice for multilingual deployments
- Sparse activation makes self-hosting more tractable on compute terms, though total memory for all experts remains high
Limitations
- We do not have a confirmed context window figure for this variant, so effective input limits must be checked per provider
- Modality support (image or audio input) is not tracked in our data for this model
- Measured time to first token of roughly 1,375 ms is not especially fast; latency-sensitive interactive apps may prefer a smaller model
- No verified public benchmark scores in our database for this specific variant, making direct quality comparisons difficult
- Total parameter footprint of ~397B means all experts must be resident in memory for self-hosting, which is demanding even though only ~17B activate per token
- Throughput and pricing vary widely between the independent providers that host Qwen models, so performance is host-dependent
Key Features
•Mixture-of-experts architecture with ~397B total parameters
•Approximately 17B active parameters per token (A17B)
•Measured throughput around 95 output tokens per second (Artificial Analysis)
•Measured time to first token around 1,375 ms (Artificial Analysis)
•Member of Alibaba's Qwen3.5 model generation
•Served through multiple third-party inference providers, enabling price and region comparison
•Multilingual orientation consistent with the Qwen family's Chinese-English focus
About Qwen3.5 397B A17B
Qwen3.5 397B A17B is a large language model released by Alibaba as part of its Qwen line. The name follows the sparse mixture-of-experts convention used across recent Qwen releases: roughly 397 billion total parameters with approximately 17 billion active on any given token (the "A17B" suffix). This places it among the larger-parameter entries in the Qwen3.5 generation, where the sparse design is intended to keep per-token compute far below what a dense model of similar total size would require.
Our measured serving figures for Qwen3.5 397B A17B come from Artificial Analysis: approximately 95 output tokens per second and a time to first token of about 1,375 ms. Those numbers reflect a specific measurement setup rather than a hard ceiling — throughput and latency for Qwen models vary considerably between hosts depending on batching, quantization, and hardware. We do not currently track a confirmed context window, modality list, or benchmark suite for this specific variant, so check the provider's documentation for the exact limits offered on a given endpoint.
In practice, models of this size and shape are typically deployed for general assistant work, coding, and reasoning-heavy tasks where a large parameter count helps but full dense inference cost is unattractive. Because Qwen releases are commonly served by multiple independent inference providers rather than a single first-party API, buyers usually have several endpoints to compare on price, throughput, and region — see the pricing table on this page for the providers we currently track.
Common Use Cases
Qwen3.5 397B A17B suits workloads that benefit from a high-capacity model but where the compute cost of a dense model at similar scale would be prohibitive: general-purpose assistants, code generation and refactoring, document analysis, and multi-step reasoning tasks. Its sparse activation profile makes it a reasonable fit for batch and throughput-oriented pipelines, where the roughly 95 tokens per second measured output rate matters more than the ~1.4 second time to first token. For highly interactive, low-latency chat interfaces, a smaller Qwen variant may be a better trade. Teams already standardized on the Qwen ecosystem — or those needing strong Chinese-language handling alongside English — often evaluate this size class against comparable mixture-of-experts models from other labs on both cost per million tokens and measured throughput.
Frequently Asked Questions
How much does Qwen3.5 397B A17B cost to use?
Pricing depends on which inference provider you use and how the endpoint is billed — per-input-token, per-output-token, and any serverless versus dedicated arrangement all differ between hosts. Because Qwen models are typically served by several independent providers, rates can vary substantially for the same model. See the pricing table on this page for the current figures from the providers we track.
What does "397B A17B" mean?
It describes a mixture-of-experts model: roughly 397 billion total parameters across all experts, with approximately 17 billion parameters activated for any single token. The practical effect is that inference compute per token is closer to a 17B-class model, while total memory requirements scale with the full 397B parameter count.
What is Qwen3.5 397B A17B best used for?
It is generally aimed at general-purpose assistant work, coding, document analysis, and multi-step reasoning where a large-capacity model is useful. Its throughput profile — around 95 output tokens per second with a time to first token near 1,375 ms in Artificial Analysis testing — leans toward batch and generation-heavy workloads rather than very low-latency interactive use.
What context window does it support?
We do not have a confirmed context window for this specific variant in our database. Because context limits for Qwen models often differ between the model's native capability and what individual hosts expose, check the documentation of the provider you plan to use.
How does it compare to smaller Qwen models?
Within the Qwen3.5 generation, this variant sits at a much larger total parameter count than the compact models, which typically translates into more capacity for knowledge-heavy and reasoning tasks. Smaller siblings generally offer lower latency and lower per-token cost, so the trade-off is capacity versus responsiveness and price. Compare the specific endpoints in the pricing table to see how each option is served.