@polyweave/qwen (1.0.8)

Published 2026-07-11 15:30:48 +00:00 by Dvorak

Set up this registry in your project .npmrc file:

@polyweave:registry=https://hub.kl1.tenere.ai/api/packages/Polyweave/npm/

To install the package using npm, run the following command:

npm install @polyweave/qwen@1.0.8

or add it to the package.json file:

"@polyweave/qwen": "1.0.8"

For more information on the npm registry, see the documentation.

@polyweave/qwen

Qwen / Alibaba adapter for Polyweave — widest catalog, multimodal, 1M context

Published as @polyweave/qwen.

Synopsis

import { createQwenAdapter } from '@polyweave/qwen';
import { createTextFrame, createFrameSink } from '@polyweave/core';

var adapter = createQwenAdapter({ apiKey: process.env.DASHSCOPE_API_KEY, model: 'qwen3-max' });

adapter.source.pipe(createFrameSink(function(frame) {
  if (frame.type === 'TEXT') console.log(frame.content.text);
}));

adapter.sink.write(createTextFrame('Explain the theory of relativity'));

Install

npm install @polyweave/qwen

Why

Qwen (Alibaba) offers the widest model catalog among OpenAI-compatible providers — from the flagship qwen3-max through multimodal qwen3.5-plus (text+image+video), coding models (qwen-coder-plus), embeddings (qwen3-embedding), and the omni model. It supports 1M context windows, thinking mode (chain-of-thought up to 81,920 tokens), and multiple deployment regions. For tasks requiring a single adapter that covers text generation, vision, embeddings, and tool calling at competitive per-token pricing, Qwen is the most versatile option.

API

`createQwenAdapter(config)`

Creates an OpenAI-compatible chat completions adapter for Alibaba Qwen / DashScope.

Parameter	Type	Required	Default	Description
`config.apiKey`	`string`	No	`process.env.QWEN_API_KEY` or `process.env.DASHSCOPE_API_KEY` or `null`	API key for DashScope authentication
`config.baseURL`	`string`	No	`https://dashscope-intl.aliyuncs.com/compatible-mode/v1`	Base URL for the OpenAI-compatible endpoint. Use a China-specific URL for domestic deployments.
`config.timeout`	`number`	No	`120000`	Request timeout in milliseconds
`config.budget`	`number`	No	`0`	Total spend limit in USD. `0` disables. When exceeded, emits `BUDGET_EXCEEDED` error and `END`.
`config.name`	`string`	No	`null`	Duplex name for topology identification and PARAM targeting

The model parameter is set via PARAM frame with key model (not in the constructor config). The initial default is qwen3-max.

Returns: IFrameDuplex — an object with the following properties:

Property	Type	Description
`source`	`IFrameSource`	Readable source emitting response frames
`sink`	`IFrameSink`	Writable sink accepting request frames
`name`	`string	null`
`setReady(key)`	`function`	Sets the API key and flushes the frame buffer. Called automatically with `config.apiKey` if provided.

Frame Protocol

Inbound frames (sink):

Frame Type	Effect
`TEXT`	Appends a user message to the conversation and sends `/chat/completions`. Emits `TEXT` (streaming or batch), `TOOL_CALL`, `RECEIPT`, `END` on response.
`IMAGE`	Appends a `content: [{ type: 'image_url', ... }, { type: 'text', ... }]` user message and sends `/chat/completions`. Requires `content.url` or `content.data`/`content.b64_json` with `content.mimeType`. Emits `TEXT`, `END`, `RECEIPT`.
`TOOL_RESULT`	Appends a tool message to the conversation with `role: 'tool'`. Does not emit a response on its own — the next `TEXT` frame will include the tool results in context.
`EMBEDDING`	Sends an embedding request to `/embeddings` using `params.embedding_model` (default `qwen3-embedding`). Emits `EMBEDDING` frame with `content.vector`, `content.dimensions`, and `content.model`, then `RECEIPT` and `END`.
`PARAM`	Sets runtime parameters. Supported keys: `model`, `max_tokens`, `temperature`, `top_p`, `stream`, `enable_thinking`, `system`, `tools`, `tool_choice`, `response_format`, `embedding_model`, `budget`, `apiKey`.
`COST_CHECK`	Enters cost-check mode. Subsequent frames are buffered and a `COST` frame is emitted with cost estimates, then `END`. If `COST_CHECK` has `maxBudget`, a `BUDGET_EXCEEDED` error is emitted when the estimate exceeds the budget.
`START`	Ignored (no-op).
`END`	Ignored (no-op). Adapter emits its own `END` per response.
`TOPOLOGY_SCAN`	Emits `TOPOLOGY_REPORT` and passes through.
`TEST_INJECT`	Injects frames from `content.frames` array.
`TEST_SNAPSHOT`	Emits a snapshot report.

Outbound frames (source):

Frame Type	When Emitted
`TEXT`	Streaming: for each content delta chunk. Batch: single frame with full text. Metadata includes `model`, `usage`, `finishReason`.
`IMAGE`	If the response contains image content (multimodal models).
`TOOL_CALL`	When the model emits tool calls (both streaming and batch). Includes `name`, `arguments` (parsed JSON), and `toolCallId`.
`EMBEDDING`	Embedding response with `content.vector`, `content.dimensions`, `content.model`.
`RECEIPT`	After each response (includes token counts and cost). Streaming: progressive receipts every 50 output tokens. Batch: single receipt at end.
`COST`	Response to `COST_CHECK`. Includes `bounds`, `accuracy`, `estimateSource`, `forwardEstimate`, `inputBounds`, `maxOutputTokens`.
`ERROR`	On failure. Code is `ADAPTER_ERROR` or `BUDGET_EXCEEDED`.
`END`	Marks end of each response stream.

Error Conditions

Condition	Frame Type	Code	Message
HTTP status >= 400 from API	`ERROR` + `END`	`ADAPTER_ERROR`	API error message from response body
Network error (connection refused, DNS, etc.)	`ERROR` + `END`	`ADAPTER_ERROR`	Node.js error message
Request timeout	`ERROR` + `END`	`ADAPTER_ERROR`	`'Timeout'`
JSON parse failure on batch response	`ERROR` + `END`	`ADAPTER_ERROR`	`'Parse: '` + error message
JSON parse failure on streaming chunk	Silently caught — individual chunks are skipped	—	—
Stream error in SSE data	`ERROR` + `END`	`ADAPTER_ERROR`	Error message from stream chunk
Budget exceeded (`totalSpent >= budgetTotal` with `budgetTotal > 0`)	`ERROR` + `END`	`BUDGET_EXCEEDED`	`'Budget exceeded'`
COST_CHECK estimate exceeds `maxBudget`	`ERROR` + `END`	`BUDGET_EXCEEDED`	`'Est cost exceeds budget'`
IMAGE frame without `url` or `data`/`b64_json`	`ERROR` + `END`	`ADAPTER_ERROR`	`'IMAGE frame requires url/data'`
Embedding API error	`ERROR` + `END`	`ADAPTER_ERROR`	API error message
No API key and no PARAM `apiKey`	Frame buffers silently until `setReady` is called	—	—

Tests

npm test

Runs node --test test/rigor.test.js test/smoke.test.js. Covers: adapter construction with and without API key, setReady availability, frame buffering before API key is set, property-based testing of adapter behavior.

Requirements

Node.js 22 or later. ESM only.

Peer dependencies: @polyweave/core, @polyweave/costs.

Caveats

Qwen uses tiered pricing by input size for some models. The international endpoint (dashscope-intl.aliyuncs.com) is used by default. For China deployment, set baseURL to the China-specific endpoint. Thinking mode is toggled via PARAM enable_thinking (boolean). The adapter buffers all frames until an API key is available — if no key is provided in config and no PARAM apiKey arrives, frames accumulate in memory indefinitely.

ID	Version
@polyweave/core	*
@polyweave/costs	*
@polyweave/errors	*

@polyweave/qwen (1.0.8)

Installation

About this package