v0.10.0

• Parse Qwen3-VL reasoning tokens correctly.
• Add Qwen3-VL-2B-Thinking-4bit and Qwen3-VL-4B-Instruct-4bit models.

v0.9.2

• End reasoning blocks when tool calls arrive
• Add Qwen3-VL models

v0.9.1

• Add 4-bit quantization of gpt-oss

Full Changelog: v0.9.0...v0.9.1

v0.9.0

• Support Python-formatted tool calls from LFM2-8B-A1B.
• Expose Python.parseFunctionCall.
• Add tools parameter to LLM.init() and convenience factory methods.

v0.8.1

• Remove responseParser argument from model factories.

v0.8.0

• Add LFM2-8B-A1B-4bit
• Remove SHLLM.memoryLimit
• Update mlx-swift-examples

v0.7.0

• Add support for GPT-OSS and improve tool calling across all models.

GPT-OSS model tool calling

let chat: [Chat.Message] = [
    .system(
        "You are a helpful assistant that can provide stock prices. When asked for a stock price, you must use the get_stock_price tool."
    ),
    .user("What is the price of AAPL?"),
]

var input = UserInput(chat: chat)

guard let llm1 = try gptOSS_20B(
    input,
    tools: [stockTool]
) else { return }

let (_, _, toolCalls) = try await llm1.result
guard let toolCall = toolCalls?.first else {
    return
}

input.appendHarmonyAssistantToolCall(toolCall)
input.appendHarmonyToolResult(["price": 123.45])

guard let llm2 = try gptOSS_20B(
    input,
    tools: [stockTool]
) else { return }

let result = try await llm2.text.result

Other model tool calling

let chat: [Chat.Message] = [
    .system(
        "You are a helpful assistant that can provide stock prices. When asked for a stock price, you must use the get_stock_price tool."
    ),
    .user("What is the price of AAPL?"),
]

var input = UserInput(chat: chat)

guard let llm1 = try qwen3MoE(
    input,
    tools: [stockTool]
) else { return }

let (_, _, toolCalls) = try await llm1.result
guard let toolCall = toolCalls?.first else {
    return
}

input.appendToolResult(["price": 123.45])

guard let llm2 = try qwen3MoE(
    input,
    tools: [stockTool]
) else { return }

let result = try await llm2.text.result

v0.6.0

• Cache most recently-used model to ensure follow-up inference is fast. This behavior can be controlled via SHLLM.isModelCacheEnabled.

v0.5.3

• Update mlx-swift-examples

v0.5.2

• Update to newest version of mlx-swift-examples
• Add C++17 language standard specification to the package configuration