MLX Patch Comparison Report

maclocal-api commit 037c657 (main) — 2026-03-01
Upstream mlx-swift-lm: pinned 2.30.3 (2c70054) · latest tag 2.30.6 · main 11968af
  pinned→latest: 21 commits · latest→main: 5 commits
Upstream mlx-swift: pin 0.30.3 · latest tag 0.30.6 · main a2f0d76
Generated: 2026-03-02
20
Patch Files
11
No Conflict
8
Low/Med Risk
1
High Risk
10
New Files

File-by-File Comparison

File Target Type maclocal-api vs Pinned vs Latest Tag vs Main Risk maclocal-api Features
Qwen3VL.swift MLXVLM MODIFIED 1787L CHANGED 1766L (+35/-14) CHANGED 1766L (+38/-17) CHANGED 1766L (+38/-17) LOW Qwen3 VL patches
Qwen3Next.swift MLXLLM NEW 723L N/A CHANGED 800L (+347/-424) CHANGED 800L (+347/-424) LOW GatedDeltaNet model architecture
GatedDelta.swift MLXLLM NEW 449L N/A N/A N/A NONE GatedDeltaNet attention layer
Qwen3_5MoE.swift MLXLLM NEW 1012L N/A N/A N/A NONE Fused QKV/GDN/gate+up projections; fusedSiluMul Metal kernel; MLXFast.rmsNorm; adaptive sanitize
DeepseekV3.swift MLXLLM MODIFIED 717L CHANGED 563L (+202/-48) CHANGED 563L (+204/-50) CHANGED 563L (+204/-50) LOW Bug fixes; Kimi K2 reuse
MiniMaxM2.swift MLXLLM NEW 362L N/A N/A N/A NONE MiniMax M2 model architecture
NemotronH.swift MLXLLM NEW 1008L N/A CHANGED 1003L (+8/-3) CHANGED 1003L (+8/-3) LOW NemotronH hybrid Mamba+Attention model
GLM4MoeLite.swift MLXLLM NEW 662L N/A N/A N/A NONE GLM-4 MoE Lite model
GLM5MoeDsa.swift MLXLLM NEW 866L N/A N/A N/A NONE GLM-5 MoE DSA model
KimiK25.swift MLXLLM NEW 113L N/A N/A N/A NONE Kimi K2.5 model architecture
LLMModelFactory.swift MLXLLM MODIFIED 584L CHANGED 554L (+31/-1) CHANGED 568L (+33/-17) CHANGED 570L (+33/-19) LOW qwen3_5_moe/kimi_k2/kimi_k25/joyai_llm_flash/glm_moe_dsa/minimax_m2; sanitizeJSON(); toolCallFormat removal
Load.swift MLXLMCommon MODIFIED 121L CHANGED 105L (+20/-4) CHANGED 105L (+20/-4) CHANGED 105L (+20/-4) NONE .jinja/.tiktoken downloads; MXFP4 fix; relaxed weight verification
Evaluate.swift MLXLMCommon MODIFIED 1451L CHANGED 1044L (+430/-23) CHANGED 1349L (+592/-490) CHANGED 1371L (+591/-511) HIGH TopK/MinP/Presence/Composite processors; seed; logprobs; AFM_PERF; clearCache; shouldQuantizeCache
Tokenizer.swift MLXLMCommon MODIFIED 429L CHANGED 157L (+272/-0) CHANGED 157L (+272/-0) CHANGED 157L (+272/-0) NONE TikToken support; .jinja fallback; O(n) window trim
Qwen3_5MoEVL.swift MLXVLM NEW 1449L N/A N/A N/A NONE Qwen3.5 MoE VLM wrapper
VLMModelFactory.swift MLXVLM MODIFIED 398L CHANGED 390L (+10/-2) CHANGED 396L (+3/-1) CHANGED 396L (+3/-1) LOW Qwen3.5 MoE VL registration
SamplerTests.swift Tests NEW 404L N/A N/A N/A NONE Unit tests for sampling processors
ToolCallFormat.swift MLXLMCommon MODIFIED 113L N/A CHANGED 120L (+11/-18) CHANGED 120L (+11/-18) MEDIUM Qwen3 model type detection (qwen3_next, qwen3_coder)
KVCache.swift MLXLMCommon MODIFIED 1610L CHANGED 1585L (+36/-11) CHANGED 1585L (+36/-11) CHANGED 1578L (+55/-23) LOW 2x growth strategy; Updatable protocol; MXFP4 mode parameter
SwitchLayers.swift MLXLMCommon MODIFIED 354L CHANGED 200L (+163/-9) CHANGED 200L (+163/-9) CHANGED 200L (+163/-9) NONE fusedSiluMul Metal kernel; SwitchGLU gate+up fusion; memory headroom check

mlx-swift Framework: 0.30.3 → 0.30.6 → main

0.30.3 → 0.30.6 (8 commits):
6ba4827 update for mlx v0.30.6 (#353) c2a5f12 Add wired memory (#352) 5a7b694 Weakly link error domains to allow building non-Metal-supporting targets (#354) 1258900 Add release workflow that builds and attaches XCFramework assets (#351) 30d0b9b Wired Memory Management System (#348) 87c9f8d Add skill.md (#350) a1bf450 adds a note to get the submodules if you're building locally (#346) f688d89 Support for building through SwiftPM on Linux (CPU only) (#304)
0.30.6 → main (4 commits):
a2f0d76 Fix dictionary key lookup in Module.update(parameters:) (#363) 2755373 Fix wired mem race condition (#358) 0cc7956 Use CXX 20, since MLX-C uses C++ 20 code (#357) d3d0a5b add gh for the release build (#356)
Generated by Scripts/gen-patch-comparison.sh — 2026-03-02
maclocal-api 037c657 · ml-explore/mlx-swift-lm · ml-explore/mlx-swift