SwiftLM b283

SwiftLM b283-00ce868

Changelog

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b275

SwiftLM b275-3990199

Merge pull request #27 from SharpAI/feature/speculative-decoding-ci

test: add speculative decoding E2E test to CI pipeline

Changelog

• ci: replace nonexistent artifact download with fresh build for eval job (92565a9)
• test: isolate speculative efficiency in separate parallel eval run (d2481d5)
• test: isolate OOM abort with 0.8B and 2B Qwen3.5 models (5a8bf24)
• test: isolate OOM abort with 0.5B and 3B Qwen2.5 models (814cdd5)
• test: remove non-streaming speculative check to prevent Abort trap: 6 (057796c)
• ci: remove matrix to keep speculative decoding test clean (2a74e67)
• ci: matrix speculative test across 4B and 9B to assert OOM behavior (5350f9b)
• ci: downgrade speculative decoding test to 4B main model for free runner (2bd2a14)
• fix: update Package.resolved to track latest mlx-swift-lm main carrying DraftModelRef (710fccd)
• docs: add Eric Lake credit to features bullet point (89b313e)
• docs: credit Eric Lake directly in SSD streaming feature intro (328c086)
• test: add speculative decoding E2E test to CI pipeline (76c68d7)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b262

SwiftLM b262-c6e6212

Merge pull request #26 from ericjlake/feat/ssd-streaming-10x

feat: 10x SSD expert streaming speedup + speculative decoding for MoE on Apple Silicon

Changelog

• fix: point mlx-swift-lm dependency at fork until SharpAI branch exists (dc40fb7)
• feat: 10x SSD expert streaming speedup + speculative decoding for MoE on Apple Silicon (cf2c4aa)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b259

SwiftLM b259-b609c8a

Merge pull request #23 from SharpAI/chore/sync-ssd-submodules

test: implement XCTSkipIf guards to bypass HFModelSearch API tests wh…

Changelog

• test: implement XCTSkipIf guards to bypass HFModelSearch API tests when GitHub Actions IP ranges trigger transient 429 rate limit responses from HuggingFace (e7f36a6)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b257

SwiftLM b257-85cfe0b

Merge pull request #22 from SharpAI/fix/ci-pipeline-stabilization

fix: stabilize SSD-streaming CI pipeline against rate limits and metallib bounds

Changelog

• test: bypass HuggingFace API network polling tests utilizing XCTSkipIf guards when GitHub Actions transient 429 rate limit triggers (82760b7)
• ci: append --disable-swift-testing flag to explicitly disable the newer testing daemon which was maliciously throwing exit code 1 when observing zero @test markers (0841251)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b253

SwiftLM b253-80c0d16

Merge pull request #20 from SharpAI/chore/sync-ssd-submodules

chore: sync ssd-prefetch and ecosystem submodules via 270GB verification

Changelog

• ci: orchestrate test environment using debug profiles recursively injecting metallib contexts adjacent to execution artifacts (d1a8308)
• ci: enforce test execution in release mode and inject MLX_METAL_PATH context to resolve metallib inference failures (167a567)
• ci: combine redundant build and e2e test github action workflows into single sequential pipeline (22be1ea)
• fix: restore unintentionally dropped SwiftBuddy and test targets from Package.swift map (e180744)
• fix: resolve subsequent GPU deprecation warning and var mutation warnings in Server and Calibrator (d864f42)
• fix: commit auto-resolved Package tree following upstream mlx-swift-lm relative path remediation (4f1cc23)
• fix: bump Package.resolved to pull newly merged ssd-prefetch API and fix compilation warnings (19f91c7)
• fix: entirely decimate rogue git submodules to match upstream SPM structure (40a0d48)
• fix: repair mlx-swift submodule path binding in .gitmodules to resolve github actions sync failure (fecf371)
• bench: wire ssd-prefetch flag to backend API and run safely on 270GB (277672b)
• chore: sync mlx-swift ecosystem submodules for ssd optimization logic (92aae79)
• docs: add PAPPS architecture research and engineering review (e8b09a3)
• chore: add WIP_REPOS README and patch profile_runner for GLM-5.1 (3224180)
• Lock MLX and SwiftLM dependencies for GLM-5.1 MoE port support (fdd343d)
• Repoint MLX packages to local glm5.1 harness (16338b7)
• bench: add GLM-270B model, memory overcommit guard, and download progress (b518551)
• test(e2e): construct full user journey harness (514ebd0)
• feat(memory): secure extraction service against hallucinated text output (65b59d1)
• ci: append swiftbuddy mempalace and lifecycle tests to build action (b9c45d8)
• refactor(mempalace): update architecture post-launch (fab4ea8)
• test(lifecycle): strict memory and device budgeting limits (f1b62c0)
• test(model-management): stabilize tests and fix regex sizes (7aeb658)
• feat(mempalace): implement Tier 6 AAAK, AutoSave, and WakeUp (Features 31-34) (a3de798)
• feat(mempalace): implement Tier 5 Temporal Knowledge Graph (Features 26-30) (605d317)
• feat(mempalace): implement Tier 3 File System Mining (Features 13, 16) (50dae71)
• feat(mempalace): implement Tier 1/2/4 parity features (45eb832)
• feat(swiftbuddy): MemPalace v1, native macOS theming, HF model management, TDD harness (63ee81f)
• feat: add GLM-5.1-4bit to benchmark suite with ssd-only restriction (c8050d5)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b222

SwiftLM b222-baebf6b

Fix HuggingFace trendingScore sort parameter

Changelog

• Fix HuggingFace trendingScore sort parameter (baebf6b)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b221

SwiftLM b221-0665bd0

Merge pull request #15 from SharpAI/feature/use-upstream-mlx-swift

Feature/use upstream mlx swift

Changelog

• fix: explicit Tokenizers dependency for MLXInferenceCore to prevent module build races (0f59dad)
• fix: bust github actions SPM cache to clear stale upstream C++ files (852fa86)
• feat: securely sparsify benchmark cloning locally to generic homesec-benchmark namespace (d5897eb)
• feat: dynamically auto-clone homesec-benchmark logic using sparse checkout if missing (9c87932)
• fix: update build.sh MLX_SRC to point to remote tracker checkouts and instantiate package resolutions securely (0753297)
• chore: wipe out lingering tracked files from former local package injection (fd107c3)
• chore: remove LocalPackages/mlx-swift proxy repo now that SharpAI remote tracks edits (50ab9b6)
• chore: align inference engine loaders and update dependency hashes (8c6dcd2)
• feat: implement streaming json output sanitization for prefixes (edba29e)
• ci: remove branch push triggers to fix PR pipeline duplication (f1bc236)
• fix(ci): Resolve mlx-swift-lm v3 API updates and update dependencies (b20f16d)
• docs(workflows): add SharpAI custom patches inventory to upstream sync workflow (357db2e)
• chore: bump mlx-swift local package reference to head (e01279a)
• chore: update mlx-swift submodule reference and package lock (c17312b)
• Remove obsolete TokenizerLoader and Downloader bridge structs (0c84bc5)
• chore(submodules): bump mlx-swift to fix C++20 build and signatures (5718e74)
• fix(deps): use remote URL dependencies for mlx-swift and mlx-swift-lm (98e43d8)
• chore: properly register SharpAI/mlx-swift as an isolated git submodule (e7f4224)
• docs: expand upstream sync flow to include verification, integration, and validation steps (1395456)
• docs: add workflow skill for MLX upstream synchronization and SSD-streaming triage (0fdfea7)
• chore: migrate mlx-swift from hardcoded source partition to git submodule (9bc651a)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b198

SwiftLM b198-29f11a6

Revert "fix(deps): use remote URL dependencies for mlx-swift and mlx-swift-lm"

This reverts commit 7f62ac9.

Changelog

• Revert "fix(deps): use remote URL dependencies for mlx-swift and mlx-swift-lm" (29f11a6)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.

SwiftLM b197

SwiftLM b197-7f62ac9

fix(deps): use remote URL dependencies for mlx-swift and mlx-swift-lm

Changelog

• fix(deps): use remote URL dependencies for mlx-swift and mlx-swift-lm (7f62ac9)

Download

• macOS Apple Silicon (arm64)

Quick Start

Please refer to the Getting Started section in the README for full installation and usage instructions.

Note: mlx.metallib is bundled in this archive. Keep it in the same directory as the SwiftLM binary — Metal GPU compute will fail if it is missing.