Skip to content

LocalLM Lab v2.1.0-GA

Latest
Latest

Choose a tag to compare

View all tags
@ancientcomputing ancientcomputing released this 15 Oct 17:10
· 14 commits to main since this release
v2.1.0-GA
1540e2f
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 2.1.0-GA

2.1 comes out of building LocalLM Lab and JevDK for the Mac App Store and submitting both, running two real apps on the SDK through App Store review.

SDK: fixes and a newer MLX runtime, no API change

The public surface is identical to 2.0.0.

  • Apps that embed LocalLMLabSDKInference now pass App Store validation. Its bundle ID was mlx-swift_Cmlx, which App Store Connect rejects (90277). It's now lab.locallm.sdk.inference, and MLX still finds its Metal library. No more renaming it or copying mlx-swift_Cmlx.bundle by hand.
  • Newer MLX runtime in LocalLMLabSDKInference: mlx-swift-lm 30d755c (upstream 8ec6a2c, 2026-10-15, plus one fix of ours), mlx-swift 0.32.3. Gemma 4 keeps array and object tool arguments intact, gpt-oss tool calls work (Harmony format), Muse Glimmer loads and calls tools, and a Qwen3 call that opens a <think> block it never closes now runs instead of coming back as text. A malformed tool call now fails the turn with LocalLMLabError.generation instead of coming back as the answer's text. Details in the CHANGELOG.
  • More local models can use tools, and model setup is sturdier. Before each turn the SDK fits the conversation to what the model's own chat template can show: GLM-4-0414 now sees tool results, Llama 3.1 and 3.2 handle a turn with several tool calls, and Mistral-Small-3.2-24B (mlx-community conversions) calls tools through a built-in template in Mistral's format. Models whose Hugging Face repo has a subfolder (every mlx-community OptiQ conversion) now download, and a model's capability check can no longer hang adding it.
  • Private Cloud Compute needed no SDK change. To use it, your app needs Apple's managed PCC entitlement in its provisioning profile, and must run PCC in its main executable (not a helper process).

LocalLM Lab app

  • Mac App Store version (not yet reviewed): Private Cloud Compute will be in the Mac App Store version, which is sandboxed. Playground Benchmark works with PCC: TTFT, total time, token counts, and an output rate. That rate includes the network round-trip, so it isn't your Mac's speed.
  • This DMG: no Private Cloud Compute. It needs Apple's managed entitlement, which only the App Store version carries. The greyed-out PCC row in AI Models is gone.
  • Both versions: Benchmark now reports token counts and a wall-clock output rate for Apple's on-device model.

Checksums

Asset SHA-256
LocalLMLabSDKCore-2.1.0-GA.xcframework.zip 8d541e06f81d1586c959d6159a259c759853c88b2c68385ab28de493af6c5776
LocalLMLabSDKClaude-2.1.0-GA.xcframework.zip f54472c25063f7c6d6d5a49a91992beb1300d9b6d986ee83d41e682ac2a2c172
LocalLMLabSDKInference-2.1.0-GA.xcframework.zip 77d3d1a1bae9df6de5955dfa945242c177cb129b4896c95befb81db6c923656f
LocalLMLabSDKRemote-2.1.0-GA.xcframework.zip 347257172a6da7104b7081602e9dbfc853ed5d038b5e30831aad8dfa2753ae92
LocalLM.Lab-2.1.0-GA-arm64.dmg (the app) 5402459733d36f2cf982afc018cd145d0e1da3ddf0f41c2c5bae5a3019d59e18
localai-toolkit-2.1.0-GA-arm64.zip (the localai-cli toolkit) 1996337f03f7a7325da066f30bacc2cbfa296730ad1e0aaa934259b09d4ab1ba

The DMG's .sha256 names the file LocalLM Lab-2.1.0-GA-arm64.dmg (with a space), so compare with shasum -a 256 LocalLM.Lab-2.1.0-GA-arm64.dmg rather than shasum -c.

Assets 14
Loading