LongCat-Flash-Chat Official Model Card: MoE Architecture, Benchmarks, and Tool Capabilities

One-sentence takeaway

LongCat-Flash-Chat is a non-thinking MoE with 560B total parameters and about 27B active on average. The official table shows strong tool calling and instruction following and moderately strong coding performance, while long-context graph traversal and privacy and safety metrics still require separate validation.

Use cases

• Good for: Non-thinking dialogue, tool calling, multi-step Agents, instruction execution, code generation/explanation, low-latency large-model inference, and 128K long-context tasks.

• Not for: Deploying in sensitive production systems based only on the official table; long-context retrieval, privacy, safety, and cross-language scenarios in particular require local retesting.

• Applicable model versions: The open-source meituan-longcat/LongCat-Flash-Chat; later API versions may differ from the open-source weights.

• Applicable clients, Agents, or APIs: Hugging Face weights, SGLang/vLLM, and the local chat template; the historical LongCat API is also subject to the retirement timeline.

• Recommended inference tier and parameters: The model card labels it non-thinking; no standardized temperature/top-p/sampling configuration is disclosed.

Test environment

• Evaluator: Meituan LongCat Team.

• Comparison models: DeepSeek V3.1, Qwen3 MoE-2507, Kimi-K2, GPT-4.1, Claude 4 Sonnet, and Gemini 2.5 Flash.

• Model scale: LongCat-Flash has 560B total parameters and 27B active on average (a dynamic range of 18.6B–31.3B).

• Inference-efficiency claim: The official claim is that inference exceeds 100 TPS on H800; the model card gives a 128K context, while an API update in 2025-12 also claims support for 256K, so the versions must be distinguished.

Input/configuration

• Mode: non-thinking foundation model.

• Tools: The official format uses <longcat_tool_call> XML tool calls, with the function name and arguments placed as JSON inside the tag.

• Deployment: The repository provides basic SGLang/vLLM adapters and a deployment guide; complete hardware, quantization, batch, and sampling configurations are not disclosed.

• Evaluation note: The comparison table notes that some external figures marked with * come from other public reports, and that several comparison models were evaluated in non-thinking mode.

Results data

The LongCat-Flash figures published in the model card (the parentheses indicate the metric definition):

Conclusions

The official table supports the view that LongCat-Flash is competitive on tool use, instruction following, dialogue, and foundational coding tasks; τ²-Bench telecom 73.68, IFEval 89.65, and SWE-Bench Verified 60.4 are useful task signals. GraphWalks-128k at just 51.05 shows that “128K support” cannot substitute for long-context graph traversal performance; safety and privacy scores should also be considered alongside real policy testing.

Limitations

• The figures come from the official model card; the full prompts, hardware, sampling configuration, and independent verification are not disclosed.

• The model card and API Change Log use two context figures, 128K and 256K, for different versions; they must not be combined into one snapshot.

• Some comparison models use non-thinking mode, and some figures marked with asterisks come from other reports; cross-column comparisons require caution.

• The model card explicitly warns that downstream applications should assess accuracy, safety, and fairness themselves; official scores do not authorize deployment in sensitive scenarios.

Reproduction steps

1. Fix the open-source commit, tokenizer, inference engine, quantization, context, temperature, and batch configuration.

2. Use the official chat template to test tool-free, multi-turn, and tool-calling flows separately, saving the original messages and XML parsing logs.

3. Retest in general, instruction, math, coding, Agent, long-context, and safety buckets, recording tokens, throughput, latency, and recovery from failures.

4. Report the historical API and local weights separately, labeling the 128K/256K version and the model's retirement status.

Source excerpts or observations (for compliant short quotations only)

• The model card defines LongCat-Flash as a non-thinking model and gives its dynamic activation range as 18.6B–31.3B.

• The official Agent evaluation covers telecom, airline, and retail variants of τ²-Bench, showing that tool-task performance varies by scenario.