edgewise/ai-sdk

Vercel AI SDK

edgewise/ai-sdk wraps Edgewise models as AI SDK providers (specification v4), so streamText, generateText, embed and tool calling in the ai package run on the device.

bash

bun add edgewise ai @ai-sdk/provider

Language models#

import { streamText } from 'ai';
import { edgewise } from 'edgewise/ai-sdk';

const result = streamText({ model: edgewise('text:default'), prompt: 'Write a haiku about the sea.' });
for await (const delta of result.textStream) process.stdout.write(delta);

Images and audio in AI SDK file parts are passed to models that accept them. AI SDK tools work: Edgewise returns the tool calls and the AI SDK runs them.

Embeddings#

import { embedMany } from 'ai';
import { edgewiseEmbeddingModel } from 'edgewise/ai-sdk';

const { embeddings } = await embedMany({ model: edgewiseEmbeddingModel('embed:default', { dimensions: 256 }), values: docs });

Evaluation models#

edgewiseEvaluationModel() implements the evaluation model interface, so the same questions can run on a local encoder or, with toAiSdkQuestions(), on a cloud judge.

Mixing local and cloud#

import { openai } from '@ai-sdk/openai';

const model = (await capabilities()).tier === 'cpu' ? openai('gpt-5-mini') : edgewise('text:default');

edgewise(id) picks the right wrapper from the model's verb: generate, embed or evaluate.