Skip to content

LocalLM Lab v2.2.0-GA

Choose a tag to compare

View all tags
@ancientcomputing ancientcomputing released this 20 Oct 19:57
· 21 commits to main since this release
v2.2.0-GA
48279a9
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LocalLM Lab 2.2.0-GA

2.2 brings NVIDIA's Nemotron 3.5 to LocalLM Lab and the SDK: the Lightning, Bolt and Super models run locally, with tools.

SDK: Nemotron 3.5, no API change

The public surface is identical to 2.1.0.

  • Nemotron 3.5 runs in LocalLMLabSDKInference. The MLX runtime now reads Nemotron 3.5's configuration and architecture, including the Bolt models (nemotron_h_puzzle). Tested with tools, including the GitHub MCP server: Lightning 30B-A3B (mlx-community 4-bit and 8-bit), Bolt 74B-A9B 3-bit and Super-VL 120B-A12B 3-bit (rapid-mlx; text only). Bolt builds that take images (nemotron_h_omni) don't load yet. Which conversions work: MLX model list.
  • effort: .off turns Nemotron's thinking off. Before, it was ignored.
  • Reasoning stays out of the answer. Nemotron 3.5's and Qwen3.5's chat templates open <think> in the prompt, so their replies started mid-reasoning and the reasoning landed in the answer. The SDK now restores the opening tag, so hostTranscript splits the reasoning from the reply with no model-specific code.
  • MLX runtime pin: mlx-swift-lm dea880a, which is 2.1.0's runtime (30d755c) plus the two Nemotron 3.5 commits. Details in the CHANGELOG.

LocalLM Lab app

  • Nemotron 3.5 models in AI Models, with thinking on or off, through the same runtime.
  • Help → Privacy Policy, and a Privacy Policy link in the About panel.
  • Mac App Store version: a Report… button on Prompt Playground's output, which opens a pre-filled email about a response. DMG-only developer settings are left out of that build.

Known issues, including Qwen3.5-9B's tool calls often being rejected as malformed: known issues.

Checksums

Asset SHA-256
LocalLMLabSDKCore-2.2.0-GA.xcframework.zip 5efc40cae9f8a1bf53bfe03195c5365398bc9501b3e6b0feee3e5c67385537f4
LocalLMLabSDKClaude-2.2.0-GA.xcframework.zip 07e2ff01accf582f71fecab254bdae2e8fb46acc8296ef6ab1c8d84cd9e97a04
LocalLMLabSDKInference-2.2.0-GA.xcframework.zip afd427fb741a820f55bc0a384850ae6404608f5ad73d185a583a89e8f0d61a61
LocalLMLabSDKRemote-2.2.0-GA.xcframework.zip 1c9af025c700bc0c41c80387c1d6a59ba2f2ab6203eb8ffd8548d61c6ac1fe38
LocalLM.Lab-2.2.0-GA-arm64.dmg (the app) 8fb39e28e75cb8add1d0ed40129237806400240d1be773535d4cca8c81a1ab73
localai-toolkit-2.2.0-GA-arm64.zip (the localai-cli toolkit) 47e485b9a07f4044d31083747ac366ee84c30536995b61f510cccaa98567cd23

The DMG's .sha256 names the file LocalLM Lab-2.2.0-GA-arm64.dmg (with a space), so compare with shasum -a 256 LocalLM.Lab-2.2.0-GA-arm64.dmg rather than shasum -c.

Assets 14
Loading