Views
No views yet
.mlpackage derived from Qwen/Qwen2.5-0.5B-Instruct, converted for use with Apple's AnyLanguageModel Swift framework and swift-transformers ≥ 1.0.inputIds / attentionMask convention required by swift-transformers 1.x LanguageModel. Drop-in compatible with CoreMLLanguageModel(url:computeUnits:chatTemplateHandler:).| Property | Value |
|---|---|
| Base model | Qwen/Qwen2.5-0.5B-Instruct |
| Precision | Float16 |
| Context window | 512 tokens (fixed shape) |
Input: inputIds | Int32 [1, 512] |
Input: attentionMask | Int32 [1, 512] |
Output: logits | Float16 [1, vocab_size] |
| Min deployment | iOS 18 / macOS 15 |
| Compute | .all (ANE + GPU + CPU) |
| Format | .mlpackage (mlprogram, compiled on first load) |
AnyLanguageModel.1import AnyLanguageModel
2
3let modelURL: URL = // path to this .mlpackage on disk
4let lm = try await CoreMLLanguageModel(
5 url: modelURL,
6 computeUnits: .all,
7 chatTemplateHandler: { instructions, prompt in
8 // Qwen 2.5 uses ChatML format; tokenizer.json Jinja template applies special tokens
9 var messages: [Message] = []
10 if let system = instructions?.description, !system.isEmpty {
11 messages.append(["role": "system", "content": system])
12 }
13 messages.append(["role": "user", "content": prompt.description])
14 return messages
15 }
16)
17let session = LanguageModelSession(model: lm, instructions: "You are a helpful assistant.")
18let response = try await session.respond(to: "Improve this text: ...")
19print(response.content)tokenizer.json, tokenizer_config.json, config.json) are bundled alongside the .mlpackage in this repo. Keep them as siblings on disk — swift-transformers resolves the chat template from tokenizer_config.json at runtime.Qwen/Qwen2.5-0.5B-Instruct by Qwen Team / Alibaba Cloud.