Your Mac is the AI lab now.

Serve, train, quantize, compose and dissect open LLMs, Meta Muse Glimmer 30B, 44.1 kHz neural speech, OCR and image synthesis on Apple Silicon. One Swift 6.4 binary speaks OpenAI API, Anthropic API and MCP with a native macOS 27 Liquid Glass app on top. Zero Python at runtime.

brew tap goodmai/basalt https://github.com/goodmai/basalt
brew trust goodmai/basalt && brew install goodmai/basalt/basalt
basalt compose up 24gb-voice-zonos
ANTHROPIC_BASE_URL=http://localhost:8080 ANTHROPIC_AUTH_TOKEN=local claude
APPLE SILICON UNIFIED MEMORY · ZERO-COPY Metal + ANE active
Basalt architecture: Metal GPU, Neural Engine, POSIX AIO SSD expert streaming and ZONOS3 44.1 kHz audio bus Metal GPU MLX · QTIP · ATEM Neural Engine Laya 11ms · Kokoro BASALT RawSpan POSIX AIO SSD 37.7 GB MoE · 0 Swap ZONOS3 44.1 kHz
67.02 tok/s Flint 40B-A2.5B (oQ3e) in 17.6 GB
RTF 0.23 · 4.4× ZONOS3-Mini 44.1 kHz · 43 languages
8.77 tok/s · 0 Swap 37.7 GB MoE streamed from SSD
84 Profiles 84 stacks for 16–256 GB Macs
FIVE SYSTEMS IN ONE SWIFT 6.4 BINARY

Inference, quantization, multimodal speech & vision, Compose and Model X-Ray

Every kernel and decode path is verified against reference logits (GoldenLogitsTests) and measured on Apple Silicon.

High-Speed Inference Behind Familiar APIs

Connect Claude Code, Cursor, OpenAI SDKs, OpenCode and aider in one line. Flint-40B-A2.5B decodes at 67.02 tok/s, Meta Muse Glimmer 30B streams Onyx reasoning & ATEM tools, Bonsai 2 27B fits in 7.4 GB, and 37.7 GB MoE checkpoints stream from SSD with 0 swap.

OpenAI + Anthropic + MCP Prefix-Tree KV Cache polar4 / turbo4 KV

QTIP Trellis, PV-Tuning, GSQ-RCO & PHM / OFT Adapters

Bitshift trellis QTIP 2.5 bpw fits Qwen3.8-27B in 10.0 GB and beats 3-bit affine on 16k perplexity (4.621 vs 4.702). Train LoRA, hypercomplex PHM and GPU-pure OFT adapters directly on streamed Hugging Face corpora.

QTIP 2.0–3.0 bpw Two-Stage PV-Tune Native GGUF IQMetal

ZONOS3-Mini 44.1 kHz Neural TTS & Duplex Speech

Native 44.1 kHz speech synthesis across 43 languages and 27 voices (RTF 0.228–0.263, cross-lingual CFG 1.5) via Zyphra ZONOS3-Mini, alongside Kokoro-82M (ANE+MLX), Qwen3-TTS, CosyVoice3, Chatterbox and Nemotron ASR.

/v1/audio/speech pipeline.speak ANE Fallback

Native MLX Image Studio & Vision OCR

Run FLUX.2/3-klein, Z-Image-Turbo, Iris-3B, SANA-1.5 and Step1X-Edit locally with shared text encoders, TeaCache and live TAEF SSE previews, paired with DeepSeek-OCR-2 and PaddleOCR-VL-1.5 document perception.

/v1/images/* TeaCache + Judge DeepSeek-OCR-2

Model X-Ray & Zero-Token Belnap Epistemic Gate

Inspect Marchenko–Pastur spectra, per-head logit lens, complex MVN phasor probes, Fisher–Rao geodesic drift and four-valued Belnap evidence states on the SIXTEEN_3 trilattice before decoding a single token.

SIXTEEN_3 Trilattice Reversible Overlays OpenMetrics ECE

Declarative Benchmarks & 10 Leaderboard Exporters

Declare throughput, perplexity and accuracy gates in YAML: basalt compose bench runs 17 suite adapters and exports verified bundles for ARC Prize, SWE-bench, Terminal-Bench, GAIA, BFCL and LiveCodeBench.

compose bench ARC-AGI pass@2 Lunar Ladder
DOCKER-COMPOSE FOR MODELS · 7 RAM TIERS × 12 ARCHETYPES

84 Hardware-Calibrated Compose Profiles for Every Mac

Pick your Mac unified memory tier: Basalt pins the 4 primary recommended stacks (★) and verifies resident memory budgets before loading a single byte.

★ 24gb-daily-chat ~7.6 GBgoodmai/gemma-4-12B-it-qat-4bit + Vision + ASR
★ 24gb-coding-agent ~11.8 GBQwen3.8-27B-QTIP + small-M kernel + MCP
★ 24gb-voice-zonos ~9.9 GBgemma-4-12B + ZONOS3-Mini 44.1 kHz + Kokoro ANE
★ 24gb-multimodal-desk ~10.0 GBPaddleOCR-VL + System One + ZONOS3 44.1 kHz
24gb-moe-frontier ~17.6 GBFlint-40B-A2.5B (67 tok/s) / Qwen3.6-35B SSD
24gb-document-ocr ~9.8 GBDeepSeek-OCR-2 + Belnap epistemic gate
basalt compose plan 24gb-voice-zonos
basalt compose up 24gb-voice-zonos
basalt compose up 24gb-coding-agent
NATIVE SWIFTUI DESKTOP APP · MACOS 27 LIQUID GLASS

Basalt.app — RAM-aware Compose Profiles (⌥⌘C), Duplex Voice & 2K Canvas

257 ms warm window, 0 idle CPU wakeups, 7.4 ms p95 per streamed token in a 2,000-message chat, ⌥Space Quick Ask above fullscreen apps, and one-click basalt:// deep links from Hugging Face.

Compose Profiles window in Basalt.app (dark macOS 27 Liquid Glass theme) showing detected 24 GB unified memory, starred recommended profiles and 24gb-voice-zonos YAML configuration Compose Profiles window in Basalt.app (light macOS 27 Liquid Glass theme) showing detected 24 GB unified memory, starred recommended profiles and 24gb-voice-zonos YAML configuration
MEASURED ON REAL HARDWARE (APPLE M4 PRO 24 GB)

Numbers Backed by Logs and Golden Logit Fixtures

Every benchmark claim links directly to its pull request, golden fixture and runner log in goodmai/basalt.

3.45–3.62×
Code Edit Speedup on Gemma 4 12B
Exact staged KV-round prompt lookup (18.7 → 64.6 tok/s) with identical greedy output (#265).
−72.1% PPL
QTIP Bitshift Trellis at 2.0–2.5 bpw
Qwen3.8-27B in 8.48 GB cuts 2-bit PPL from 19.15 to 5.35, and 2.5 bpw beats 3-bit affine (#494).
98.9% · −77%
Local System One Triage Judge
OpenJev-Flash-9B-4bit beside Gemma 4 12B saves 77% of downstream LLM tokens in 15.3 GB (#465).
36 s → 1 s
Second Agent Turn Latency
Prefix-tree prompt cache with recurrent hybrid anchors on a 12.6k-token agent prompt (#82, #219).
GOODMAI HUGGING FACE COLLECTION · 1-CLICK LAUNCH

42+ Verified Checkpoints Ready to Launch in One Click

Every checkpoint in huggingface.co/goodmai ships with a one-click Basalt.app deep link and verified CLI command.

goodmai/gemma-4-12B-it-qat-4bit

Daily-driver 12B QAT multimodal chat and vision model (~7.6 GB on 24 GB Mac).

goodmai/Qwen3.8-27B-QTIP-2.5bpw-MLX

27B in 10.0 GB: 2.5 bpw QTIP trellis beats 3-bit MLX affine on 16k perplexity (4.621).

goodmai/Qwen3.6-35B-A3B-8bit

35B MoE (37.7 GB) streamed from SSD via POSIX AIO at 8.77 tok/s on 24 GB with 0 swap.

Zyphra/ZONOS3-Mini-hf

Native 44.1 kHz neural TTS across 43 languages & 27 voices (RTF 0.23–0.26 in 2.24 GB).

goodmai/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed

Hadamard ternary 27B in 7.4 GB: 27.1 tok/s with MTP head and up to 38 tok/s on code.

goodmai/DeepSeek-OCR-2-8bit

Learned-query vision encoder with official 768 tiling for dense PDFs (/v1/pipeline).

mlx-community/Muse-Glimmer-30B-4bit

Meta Muse Glimmer 30B multimodal agent (131k context, Onyx to=self reasoning & ATEM tool calling).

🤗 Browse all 42+ models in the Hugging Face Collection → Bosalt Org on HF →
COMPLETE ENGINEERING DOCUMENTATION (ENGLISH)

Guides, API References, Architecture & FAQ

Frequently Asked Questions

What Mac hardware and OS are required for Basalt v0.11.0?
Any Apple Silicon Mac (M1–M4, 16 GB to 256+ GB unified memory) running macOS 27+ with Xcode 27.0+ (Swift 6.4 + MetalToolchain when building from source). Zero Python, Conda or Docker at runtime.
Do my documents, audio recordings, or prompts ever leave my Mac?
Never by default: Basalt runs 100% locally offline. Even when you opt into budgeted cloud escalation in a Compose file, perceived document and audio text defaults to data: no-perceived and never leaves your Mac.
How does automatic RAM recommendation work across the 84 Compose profiles?
When opening Compose Profiles (⌥⌘C) in Basalt.app, ComposeMemoryTier.detect reads physical unified memory and pins the 4 starred primary profiles for your tier at the top while keeping all 84 profiles searchable and selectable.

⚡ Basalt v0.11.0 Quickstart

Install the CLI via Homebrew or build the native Basalt.app bundle from source:

brew tap goodmai/basalt https://github.com/goodmai/basalt
brew trust goodmai/basalt
brew install goodmai/basalt/basalt
basalt onboard
git clone https://github.com/goodmai/basalt.git && cd basalt
./scripts/bundle_app.swift
open .build/Basalt.app
▶ Hugging Face Launcher