Qwen3.5-397B-A17B: specs, benchmarks, and how to run it on Lyceum

Get started: call Qwen3.5-397B-A17B on Lyceum

You can access Qwen3.5-397B-A17B through Lyceum Technology's Serverless Inference API. Switch to Lyceum by updating your base URL and API key. The API is fully OpenAI-compatible, requiring no other code changes.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="Qwen/Qwen3.5-397B-A17B",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Qwen3.5-397B-A17B

Serverless Inference has no service tiers: latency and capability follow from the model you choose, and Qwen3.5-397B-A17B is built for high-capability workloads. The pricing is $0.60 per million input tokens and $3.60 per million output tokens. This endpoint is globally hosted rather than EU-hosted, and global models never receive your traffic unless you explicitly select them. Billing is strictly pay-per-token with no minimum commitments or idle compute costs.

For engineering teams migrating from hyperscaler environments, this drop-in compatibility eliminates the need to rewrite application logic. You can route traffic to Qwen3.5-397B-A17B immediately using standard chat completion parameters like temperature and top_p.

Self-hosting Qwen3.5-397B-A17B requires significant infrastructure, typically an 8-GPU node of NVIDIA HGX B200 or H100 class hardware, because it is a massive 397-billion parameter model. Utilizing Lyceum's managed endpoint allows you to bypass the operational overhead of provisioning hardware, tuning vLLM configurations, and managing Kubernetes clusters. You only pay for the tokens you consume, making it highly cost-effective for both bursty workloads and sustained production traffic.

What Qwen3.5-397B-A17B is good at

Native multimodal reasoning

Qwen3.5-397B-A17B is a unified vision-language foundation model, unlike previous generations that maintained separate text and vision model lines. It uses early fusion training on multimodal tokens, allowing it to process images and video natively. This architecture enables the model to excel at visual reasoning, GUI interaction, and video comprehension without relying on external vision encoders. It achieves parity with massive text-only models while offering superior spatial awareness for tasks like automated quality inspection by processing visual and textual data in a single shared representation space. Another mixture-of-experts model from the same family is covered in our notes on Qwen3-235B-A22B and how to run it.

Agentic workflows and coding

Qwen3.5-397B-A17B operates in a "thinking mode" by default, generating internal reasoning traces (<think>...</think>) before producing a final response. This chain-of-thought approach significantly boosts its performance on complex agentic tasks and software engineering benchmarks. It is highly capable of executing multi-step planning, tool calling, and autonomous coding tasks. When integrated into an application, this allows the model to self-correct and evaluate multiple solution paths before committing to an answer.

High-throughput MoE efficiency

The model utilizes a sparse Mixture-of-Experts (MoE) architecture combined with Gated Delta Networks (linear attention). Out of 512 experts, it routes each token to 10 routed experts plus one shared expert, activating only 17 billion parameters per forward pass out of its 397 billion total. This 3:1 ratio of linear attention to full attention reduces KV-cache memory requirements by approximately 4x. For infrastructure teams, this architectural efficiency translates directly into higher decoding throughput and lower cost per token.

Limitations and trade-offs

Where Qwen3.5-397B-A17B falls short

Qwen3.5-397B-A17B delivers exceptional reasoning and coding capabilities but has documented limitations regarding factual accuracy. Artificial Analysis evaluates the model on AA-Omniscience accuracy and non-hallucination rate alongside its reasoning and coding scores. For strict factual retrieval tasks without external grounding, a reasoning-first model of this size may still confidently generate incorrect information. Engineering teams deploying this model for knowledge-intensive applications must implement robust retrieval-augmented generation (RAG) pipelines to anchor the model's responses to verified data sources.

Additionally, the model's default "thinking mode" introduces a trade-off in time-to-first-token (TTFT) for the final output. Simple queries may experience higher perceived latency compared to non-reasoning models because the model generates internal reasoning tokens before answering. While this is beneficial for complex coding tasks, it is less optimal for latency-sensitive, real-time conversational applications where immediate responses are required. If your use case involves real-time voice processing or fast autocomplete features, a smaller dense model might be a better architectural choice.

Finally, the sheer size of the model (397B total parameters) means that self-hosting it requires substantial VRAM. Even with 4-bit quantization, the model weights come to roughly 200 GB and require massive memory bandwidth to serve efficiently. For teams managing their own infrastructure, this creates a high barrier to entry, making managed API access through Lyceum Technology a more practical deployment path.

Benchmarks and how it compares

Qwen3.5-397B-A17B benchmark results

Qwen3.5-397B-A17B ranks highly among open-weights models, scoring 34 on the Artificial Analysis Intelligence Index in its reasoning configuration, read 3 August 2026. It demonstrates strong performance across reasoning, coding, and knowledge benchmarks.

Qwen3.5-397B-A17B figures are from its Hugging Face model card, cited in Sources below. GLM-5 and Kimi K2.5 are external models that Lyceum does not serve; their figures are from the GLM-5 and Kimi K2.5 model cards. All three read 3 August 2026. A dash means the card does not report that benchmark.

The model's architecture allows it to punch above its weight class. In the SWE-Bench Verified evaluation, which tests a model's ability to resolve real-world GitHub issues, Qwen3.5-397B-A17B achieves a 76.4% resolution rate. This places it in the upper echelon of coding models, rivaling proprietary alternatives. Its performance on the IFEval benchmark at 92.6% indicates high reliability when adhering to strict formatting constraints, such as generating valid JSON.

Qwen3.5-397B-A17B serves as the heavy-duty reasoning engine within the broader Qwen family. Smaller dense Qwen models such as Qwen3-32B, which Lyceum also serves, offer strong coding performance for their size, but the 397B MoE model retains a decisive advantage in complex, multi-step agentic workflows and deep scientific reasoning tasks. Against the larger external models GLM-5 (744B) and Kimi K2.5 (1T), neither of which is in Lyceum's catalogue, it leads on MMLU-Pro and GPQA and trails slightly on SWE-Bench Verified, while activating far fewer parameters per token (17B against 40B and 32B respectively), making it cost-efficient for production inference.

Using it in production

Production configuration for Qwen3.5-397B-A17B

When deploying Qwen3.5-397B-A17B via Lyceum Technology, you are calling a single Serverless Inference endpoint with no tier to select. The model is designed for high-capability, complex reasoning tasks and supports a native context window of 256K (262,144 tokens), making it ideal for analyzing large codebases, processing extensive document repositories, or handling long-running agentic workflows. Keep in mind that filling the full 256K context will naturally increase the time-to-first-token.

Ensure your application is configured to handle streaming responses, as the model utilizes a "thinking mode" to generate reasoning traces. Streaming the output prevents timeout errors and improves the user experience by displaying the reasoning process (<think>...</think>) in real-time as the model works through complex problems. If your frontend does not support streaming, you must configure your HTTP client with extended timeout thresholds, as the model may spend several seconds generating hidden reasoning tokens before returning the final JSON or text payload.

At $0.60 per million input tokens and $3.60 per million output tokens, the model is highly cost-efficient for its capability class. For a realistic production workload, such as processing a 10,000-token document and generating a 1,000-token analysis, a single API call costs approximately $0.0096. This endpoint is globally hosted rather than EU-hosted, and it never receives your traffic unless you explicitly select it. This pay-per-token model scales from zero, so you incur no idle costs during periods of low traffic and get a predictable path forward for scaling AI features.

Why run Qwen3.5-397B-A17B on Lyceum

Why run Qwen3.5-397B-A17B on Lyceum

Managing the infrastructure required for a 397B-parameter MoE model is a significant operational burden for AI startups and enterprise engineering teams. Lyceum Technology simplifies this process by providing a fully managed, OpenAI-compatible API. You get the capabilities of Qwen3.5-397B-A17B as a drop-in replacement, without the overhead of provisioning 8-GPU nodes, tuning vLLM configurations, or managing Kubernetes clusters. Our platform is designed for engineers who build, billing Serverless Inference per token with no base fee so that you only pay for the tokens you actually consume.

Qwen3.5-397B-A17B is one of four models in Lyceum's catalogue that are globally hosted rather than EU-hosted, and it never receives your traffic unless you explicitly select it. Lyceum passes its structural cost advantage straight through to you: highly competitive per-token pricing, unified billing, and no egress fees. There are no idle or base charges, so you pay strictly for the tokens you consume, and you can burst to per-second dedicated GPUs whenever a workload needs more headroom.

Furthermore, Lyceum's open-stack transparency, built on vLLM, NVIDIA Dynamo, and TensorRT-LLM, ensures that you avoid the vendor lock-in associated with proprietary inference engines. By standardizing on open-source orchestration, we guarantee customer portability by design. If your scaling strategy eventually requires transitioning from our serverless inference API to your own dedicated Lyceum VMs, the underlying software stack remains consistent, eliminating costly migration engineering. Lyceum delivers the performance, transparency, and cost-efficiency required to scale AI applications in production.

Sources

[1] Qwen3.5-397B-A17B Model Card; [2] Qwen3.5 GitHub Repository; [3] Qwen3.5 Collection