# Flonno > An optimized, additionally quantized Qwen derivative served through Flonno's OpenAI-compatible API for text, image understanding, coding and tool workflows. Facts reviewed: 2026-10-17. These are published contract details, not a live health or speed measurement. ## Available API model - Name: Qwen3.8 Flash 27B - Model ID: qwen3.8-27b - Base URL: https://www.flonno.com/v1 - Input: USD 0.03 / 1,000,000 tokens - Output: USD 0.25 / 1,000,000 tokens, including billed reasoning - Billing: prepaid; token estimates exclude taxes, retries and other workflow costs - Context: 131072 tokens shared by input and output - Maximum requested output: 32768 tokens - Default output budget: 4096 tokens - Capabilities: text output, inline image understanding, streaming, tool calls, JSON mode - Research-library inclusion does not mean API availability. ## Documentation - [Service status and capacity](https://www.flonno.com/status) - [Reproducible agent API evaluation](https://www.flonno.com/research/agent-api-evaluation) - [Model, pricing and FAQ](https://www.flonno.com/models/qwen3-8-flash-27b) - [API reference](https://www.flonno.com/docs) - [Agent quickstart](https://www.flonno.com/developers/agents) - [OpenAPI specification](https://www.flonno.com/integrations/openapi.json) - [Provider model feed](https://www.flonno.com/integrations/openrouter/models) - [Integrations and SDK examples](https://www.flonno.com/integrations) - [OpenAI migration guide](https://www.flonno.com/migrate) - [Cost calculator](https://www.flonno.com/tools/api-cost-calculator) - [German Qwen API guide](https://www.flonno.com/de/qwen-api) - [Privacy and processing](https://www.flonno.com/privacy) - [Terms](https://www.flonno.com/terms) - [Structured model facts](https://www.flonno.com/models/qwen3-8-flash-27b.json) - [Full text reference](https://www.flonno.com/llms-full.txt) ## Deployment identity - Base checkpoint: Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 - Artifact: WarlaxZ/Qwen3.8-27B-NInfer-3090 @ d47f2732d369acaec76dc44228f67c72b081df2f - SHA-256: 7f2a0086a071ee932c66106e1552d118a85cfec8ddef13c67a54cd64abb39706 - Weight encoding: NInfer container v3; mixed Q4/Q5/Q6 weight encoding - Quantization: Importance-matrix calibrated grouped_search quantization; 4-bit embedding, 6-bit output head, and selected 4-bit attention/GDN output and MLP-down projections in layers 36–63. Other tensor formats vary. This is not uniform W4A16. - Runtime: NInfer 0.15.1-rtx3090; DFlash2 with proposal head; rk4v4 KV cache; vision overlay - License: Apache-2.0; [source model card](https://huggingface.co/WarlaxZ/Qwen3.8-27B-NInfer-3090/blob/d47f2732d369acaec76dc44228f67c72b081df2f/README.md) - Mixed tensor formats are not a uniform W4A16 claim. Upstream benchmark results are not Flonno API results. ## Processing and limitations Qwen3.8 Flash 27B runs on our own GPU system in Switzerland. API requests go from the Flonno gateway to Flonno's inference runtime; they are not sent to another AI API provider for model inference. The current deployment uses an RTX 3090 and has no automatic external-provider fallback. Hosting infrastructure remains part of the processing chain; authentication and payments use separate service providers described in the privacy notice. Evaluate output quality on your own tasks. This optimized derivative is not unchanged original Qwen weights. No universal task coverage, quality score, throughput guarantee or automatic recommendation by any search assistant is claimed. See the privacy notice for retention and legal details. ## Chat Completions quickstart API key required. Set FLONNO_API_KEY in your local environment; never put it in source control or a browser bundle. ```sh curl https://www.flonno.com/v1/chat/completions -H "Authorization: Bearer $FLONNO_API_KEY" -H "Content-Type: application/json" -d '{"model":"qwen3.8-27b","messages":[{"role":"user","content":"Summarize this text in one sentence: The API supports streaming and tool calls."}],"reasoning_effort":"none","max_tokens":200}' ``` ## Request handling - POST /v1/chat/completions: OpenAI-compatible chat. This does not imply support for every OpenAI API or SDK feature. - POST /v1/messages: Anthropic-compatible messages; inspect the API reference for supported fields and tool-client setup. - GET /v1/models: authenticated model availability; public facts are documentation, not a replacement for authentication. - Tool selection: use tool_choice: auto or omit it. Forced tool_choice: required is unsupported; automatic selection is not a tool-call guarantee. - stream: true returns SSE. Handle partial events and completion usage. - response_format: {"type":"json_object"} requests JSON mode. Validate required keys and value types in your application. - Send image_url message parts as PNG, JPEG or WebP base64 data URLs. Maximum four images, 2 MiB each, 4 MiB JSON request body. Remote image URLs and image generation are not supported. - Treat 401 as an authentication issue, 402 as a credit issue and 429 as a rate/concurrency signal. Check the actual error body; bound retries and avoid replaying irreversible tool actions. - Scoped keys, explicit token budgets, tool allowlists and output validation are recommended. ## Frequently asked questions ### Is Flash the original model name? Qwen3.8 Flash 27B is Flonno’s deployment name; use qwen3.8-27b in API requests. It is an optimized, additionally quantized Qwen3.8 27B derivative, not Alibaba’s separately named Qwen3.8-Flash service. Original-model benchmark scores are not deployment-specific measurements. ### What does the Flonno Qwen3.8 Flash 27B API cost? The token rates are $0.03 per million input tokens and $0.25 per million output tokens, in USD. For 10,000 requests with 1,000 input and 300 billed output tokens each, the token estimate is $1.05 before taxes, retries and other workflow costs. Billed reasoning counts toward output usage. ### Which model ID and base URL should an agent use? Use model ID qwen3.8-27b, base URL https://www.flonno.com/v1, and a scoped Flonno API key. Chat Completions uses POST /v1/chat/completions. API requests require authentication and prepaid credit. The public model library also contains research records for models that are not served by Flonno. ### What can I build with this API? Evaluate ticket classification, document-field extraction, summaries, retrieval-grounded answers, coding assistance and bounded tool workflows. The API supports streaming, tool calling, JSON mode and inline image inputs. Your application executes tools and validates results; a model-generated tool call is not an executed action. ### Is Qwen3.8 Flash 27B sufficient for every task? No model is reliable for every task. Test this deployed, additionally quantized derivative against your own accepted answers. Original-model benchmarks do not establish this deployment's accuracy. Review generated code and factual answers, validate extracted fields, and require human review for consequential decisions. ### Does Flonno operate its inference runtime? Qwen3.8 Flash 27B runs on our own GPU system in Switzerland. API requests go from the Flonno gateway to Flonno's inference runtime; they are not sent to another AI API provider for model inference. The current deployment uses an RTX 3090 and has no automatic external-provider fallback. Hosting infrastructure remains part of the processing chain; authentication and payments use separate service providers described in the privacy notice. ### Is this a local or offline API? No. The public service is a hosted HTTPS API. Your request reaches the Flonno gateway and its configured GPU runtime. Review the privacy notice and service terms for the processing path, retention policy and recipient categories before sending sensitive information. Operating a GPU runtime does not by itself establish end-to-end encryption or zero retention. ### What are the context and output limits? Input and output share a 131,072-token context ceiling. Requested output is at most 32,768 tokens, and the omitted output allowance defaults to 4,096 tokens. Reasoning defaults to medium; use reasoning_effort: none for non-thinking requests. Reasoning and visible output share the output budget.