Repository navigation
Basalt v0.11.0
brew tap goodmai/basalt https://github.com/goodmai/basalt
brew trust goodmai/basalt
brew install goodmai/basalt/basaltBuilds from source: Xcode 27.0+ (Swift 6.4) and macOS 27+, Apple Silicon.
Source tarball SHA256: 48e35f3809acfb3afafbeed0381335a4dc661abbb7c2690ced83ddc556117244
What's Changed
- chore: point formula at v0.10.0 by @github-actions[bot] in #382
- speech: Nemotron 3.5 Transcribe (nemotron_transcribe) model, /v1/audio/transcriptions, and SpeechAnalyzer fallback (#230, #222) by @goodmai in #383
- feat(cloud): segment-level speculative cascades between local and cloud models by @goodmai in #389
- feat(embeddings): bidirectional mask, float32 mean pooling, and headless LoRA (#384) by @goodmai in #390
- feat(lab): GoodViewCollator, JEPAProjector, LeJEPALoss (#385) by @goodmai in #391
- lab: train_jepa (#386), Orthogonal Procrustes (#387), matmul Quest block scoring by @goodmai in #392
- quantization: NF4 phase (c) — native nf4 mode from our mlx fork, experts included, experimental (#357) by @goodmai in #395
- lab: distill — teacher → student LoRA from an API-key teacher, completion-only loss, cached and budgeted (#393) by @goodmai in #396
- Batch 13: fused Quest kernels, MTP draft head, abliteration and GPTQ measurements, distillation fixes (#280 #209 #44 #376) by @goodmai in #398
- Batch 14: on-policy distillation, GPU/ANE contention rule, Core AI embeddings on the ANE, prefill-split spike step 1 (#394 #273 #223 #274) by @goodmai in #401
- Batch 15: System One decisions, OCR phase 0, Apple Vision OCR, combine pipeline phase 1, compute plan (#234 #400 #403 #402 #399) by @goodmai in #406
- Batch 16: PaddleOCR-VL, preemptible distill, ANE System One, pipeline audio, NF4 kernels, MLX↔Core AI hand-off (#400 #399 #234 #402 #357 #274) by @goodmai in #411
- Batch 17: DeepSeek-OCR, pipeline presets and sequential memory, sidecars by rule, context-scaling bench, per-token trace, cascade System One (#400 #402 #399 #242 #405 #407) by @goodmai in #414
- vision: ANE vision tower for PaddleOCR-VL (einsum export, CER 1.74% vs 1.90%) behind an experimental flag; plan notes from batch 17 measurements (#224 #399 #402) by @goodmai in #415
- ocr: DeepSeek-OCR no-repeat 20-gram processor, as in the official inference; gundam CER 15.6% → 3.8% (#400 follow-up) by @goodmai in #416
- compose: basalt compose plan|up — YAML model systems with local and cloud models, on-demand/sequential keep (#412 phase 1) by @goodmai in #417
- compose: answer routes (MoM), multi-cloud decide cascades, cloud budgets and the 401/429 policy (#412 phase 2) by @goodmai in #418
- compose: named pipelines, up-as-diff on a running system, compose ps/down, docs/compose.md (#412 phase 3) by @goodmai in #419
- deps: bump github.com/apple/swift-log from 1.16.2 to 1.17.0 by @dependabot[bot] in #420
- deps: bump github.com/apple/swift-nio from 2.104.0 to 2.105.0 by @dependabot[bot] in #421
- gguf: native GSQ-RCO i-quant/k-quant kernels and GGUF import, experimental (#408 phases 0 and 2) by @goodmai in #424
- compose v2 phase 4a: schema as data, exact lines, version 2 variables and profiles, paths next to the file (#423) by @goodmai in #425
- deps: bump swift-jinja from 2.5.2 to 3.1.0 (major), checked against real chat templates by @goodmai in #426
- compose v2 phase 4b-1: cloud budgets and key variables apply to a running system (#423) by @goodmai in #428
- compose v2 phase 4b-2: admission on apply, unload-first when old and new do not fit together, plan counts every pipeline (#423) by @goodmai in #429
- compose v2 phase 4b-3: the local decision model is loaded once, by serve's decision engine (#423) by @goodmai in #430
- gguf: half-precision lattice path for IQ3_S and IQ3_XXS matvec, decode +4.7% on Qwen3.8-27B (#408) by @goodmai in #431
- compose v2 phase 4b-4: the runtime runs on narrow injected services; a re-keyed cloud model uses its new key (#423) by @goodmai in #432
- gguf: IQ4_XS matvec reads each 16-byte group with one load, −14% on the kernel (#408) by @goodmai in #433
- Mac app phase 1: SwiftUI chat on a supervised basalt helper, SwiftData history, Monitor window, bundle script (#434, closes #55) by @goodmai in #435
- compose v2 4b-5: compose up changes the chat model and the local decider without a restart; /v1/systemone follows (#423 A2) by @goodmai in #436
- compose v2 4b-6: each pipeline decides with its own local model, loaded once (#423 A1) by @goodmai in #437
- compose v2 4c-1: closed model states, /v1/health/live and /ready, unset keys fail up, pid reuse, down fails on timeout (#423) by @goodmai in #438
- compose v2 4c-2: bind first and load after; ps shows loading; a second up waits and applies; up --wait (#423) by @goodmai in #439
- compose v2 4c-3: startup probes; a failed probe shows as failed and keeps ready at 503 (#423) by @goodmai in #440
- Mac app phase 2: menu bar item, quick ask on ⌥Space, Settings with Keychain cloud keys and open at login, restored selection and drafts (#434) by @goodmai in #441
- Mac app phase 3a: Models window — fit verdicts gate loading, resumable downloads, safe delete (#434) by @goodmai in #442
- Mac app phase 3b: PDF, image and audio attachments through /v1/pipeline, perceived text kept for follow-ups (#434) by @goodmai in #443
- Mac app phase 3c: inspector — per-conversation system prompt, receipt details, perceived attachments (#434) by @goodmai in #444
- Mac app phase 4a: App Intents, Ask Basalt service, Dock drop, answer notifications (#434) by @goodmai in #445
- Mac app phase 4b: model or compose system with switching, adoption by arguments, compose and cloud spend in the Monitor, cloud mark (#434) by @goodmai in #446
- compose v2 4d: compose pull, plan sized from the Hub or refused, compose diff dry run, progress events (#423) by @goodmai in #447
- decisions: Gemini Decision 1 as a cloud stage — cloud:gemini/, multimodal media on /v1/systemone, compose data: media opt-in by @goodmai in #449
- Mac app phase 5a: VoiceOver, Russian localization with a coverage test, daily update check (#434) by @goodmai in #456
- Mac app phase 5b: zip, dmg, notarization, release step and Homebrew cask when signing secrets exist (#434) by @goodmai in #457
- telemetry: trailing windows, decision aggregates and OpenMetrics at /api/v1/metrics (#450 phase 1) by @goodmai in #458
- ci: cap Swift Testing parallelism so MLX tests on the eval lock stop starving live-server tests (#427) by @goodmai in #461
- compose v2 4e-1: cascades in any local/cloud order, readout calibration from the bundle, OpenJev sizing (#423) by @goodmai in #460
- telemetry: MoE routing and expert cache, per-token confidence shape, telemetry_query MCP tool, Monitor sections (#450 phases 2–4) by @goodmai in #459
- compose v2 4e-2: Jev LoRA applied unmerged on the stored base (4-bit included), Qwen3.5 text bases, sizing as base plus adapter (#423 J2 J3) by @goodmai in #462
- app: measure the #434 budgets — launch, idle wakeups, memory, 60 fps streaming, quick ask latency by @goodmai in #463
- jev: real-model gate for the unmerged LoRA — OpenJev-0.6B on bf16, 8-bit and 4-bit bases (#423 J2) by @goodmai in #464
- jev: a 9B judge beside Gemma 4 12B on 24 GB — 4-bit fits, 8-bit swaps (#423 4e gate, #234) by @goodmai in #465
- nf4: fair comparison from bf16 on Gemma 4 12B — NO-GO on quality (perplexity +29%), the kernel is correct (#357) by @goodmai in #466
- ane: #274 step 3a — the ANE weight format for the prefill FFN split on Gemma 4 12B: int8 per channel is the only one accurate and fast by @goodmai in #467
- ane: #274 step 3b-2 — SplitFeedForward: a dense layer's MLP as the GPU's columns plus a Neural Engine share, through Gemma 4's async prefill by @goodmai in #468
- docs: lead with what is measured — golden-tested models on the first screen, every number linked, a README model check in CI (#404) by @goodmai in #470
- kv: experimental --kv-bits turbo3 | turbo4 | turbo8v4 — the fork's TurboQuant cache on full-attention layers, with the measured trade-off (#471) by @goodmai in #472
- kv: experimental --kv-bits polar4 — polar RoPE-pair keys in the zoned middle (#351) by @goodmai in #473
- vision: DeepSeek-OCR-2 port — Qwen2 encoder with learned queries, official 768 tiling, processor by model_type (#413) by @goodmai in #476
- moe: expert streaming decode 2.3–2.6x — per-layer expert slots read in place with batched AIO (#225 phase 2) by @goodmai in #477
- moe: Qwen3.6-35B-A3B-8bit acceptance (8.77 tok/s, 0 swap), prefill warm-start, and LIO_NOWAIT next-layer prefetch by @goodmai in #478
- gui: macOS 27 Liquid Glass surface architecture & deterministic @2x visual/MP4 verification framework by @goodmai in #479
- routing: ordered LAN-before-cloud fallback tiers (#142) by @goodmai in #480
- compose v2 4e: on-demand deciders with idle unload and disk-only startup probe (#423 J5) by @goodmai in #481
- docs: generate static benchmark page and review-gated HF collection manifest (#404) by @goodmai in #482
- quantization: add --experimental-pv-tune to models convert (#410) by @goodmai in #486
- harness/skills: add model-surgery skill and harden weight-edit rules (#483) by @goodmai in #487
- mcp: configurable per-tool and per-group enable/disable toggles (#484) by @goodmai in #488
- compose: support experimental flags and mcp options with live reload (#485) by @goodmai in #489
- quantization: QTIP Phase 0 bitshift trellis converter and native loader (#409) by @goodmai in #494
- docs, app: add goodmai HF launcher buttons, basalt:// deep link, and 0.8 README refresh by @goodmai in #495
- arch: Swift 6.4 zero-hatch, Span/RawSpan, unboxed sampling, SSE fast writer, and fused confidence (#490) by @goodmai in #496
- arch: Swift 6.4 yyjson/RawSpan fast reader, generic orchestrator, de-boxed KV & fused QTIP/RHT kernels (#490, #491, #409) by @goodmai in #497
- feat(telemetry): X-ray telemetry value core, cache & calibration ECE, SwiftMetricsSink, and /metrics (#450) by @goodmai in #498
- feat(inference,core): zero-copy unified memory bridge (#490) and specialized IQ Metal kernels + RCO allocator (#408) by @goodmai in #499
- feat(systemone): zero-token Belnap FOUR epistemic gate (#452) and multi-head decision banks (#455) by @goodmai in #500
- feat(adapters): implement PHM and Compacter hypercomplex adapters (#451) by @goodmai in #501
- feat(inference,core): zero-copy spans, ownership containers, and fused decode/MoE kernels (#490) by @goodmai in #502
- feat(xray,harness): add T-norm satisfaction scores, Fisher-Rao drift, and geodesic steering (#361, #362) by @goodmai in #503
- feat(systemone,cli): add teacher-guided readout calibration and calibrate-decide (#448) by @goodmai in #504
- feat(bench): compose v2 declarative benchmarks, BasaltBench adapters, and top-10 submission exporters (#492) by @goodmai in #505
- feat(speech,server,core): add TTS synthesis plane, ANE fallback router, and /v1/audio/speech (#493) by @goodmai in #506
- feat(xray): add four-valued Belnap evidence state, refutation channel, and SIXTEEN_3 trilattice (#360) by @goodmai in #507
- feat(adapters): add GPU-pure Cayley and matrix-exp OFT adapters (#363) by @goodmai in #508
- feat(xray): add complex MVN phasor probe and FHRR binding (#365) by @goodmai in #509
- feat(train,compose): stream HF Nemotron corpora into LoRA/PHM/OFT adapters & compose multi-adapter routing (#510) by @goodmai in #511
- feat(decisions): complete #234 decide preset sizing, Laya ANE exporter, and self-calibration tests by @goodmai in #512
- feat(speech): add Kokoro-82M (ANE+MLX), Qwen3-TTS, Fun-CosyVoice3, and Chatterbox neural synthesizers (#493) by @goodmai in #513
- feat(app): complete #434 duplex voice, server-side hub boundary, hot compose switch, markdown blocks, and basalt app by @goodmai in #514
- feat(quant): add specialized IQMetalKernels, RCOSensitivityAllocator, mixed-rate QTIP budget, and two-stage QTIP PVTuner (#408, #409) by @goodmai in #515
- feat(inference,cli): graduate ANE/MoE streaming flags, retire SplitFeedForward, and add BatchQuantizedKVCache (#226, #243) by @goodmai in #516
- feat(image): native MLX image synthesis & editing, Compose imagine:, and Basalt.app canvas (#517) by @goodmai in #519
- feat(inference): add Recursive Flint 40B-A2.5B, Ricursive Ripple-65B, and NVIDIA Nemotron-3.5-Lightning (#518, #520, #521) by @goodmai in #522
- deps(inference): retire goodmai/mlx-swift fork via native NF4Metal and drop SplitFeedForwardPrefill (#275, #102) by @goodmai in #523
- deps(inference): localize regex grammar & ModelContext slots out of mlx-swift-lm (#275) by @goodmai in #524
- feat(speech,compose): support Zyphra ZONOS3-Mini 44.1 kHz TTS and Compose speak: (#525) by @goodmai in #527
- feat(compose,app): 84 hardware-tiered Compose profiles and GUI selector with RAM recommendation (#526) by @goodmai in #528
- release: 0.11.0 by @goodmai in #529
- fix(tests,ci): split synthetic sk-ant token in RemoteDecisionEngineTests and align security_audit.sh with release.yml by @goodmai in #530
Full Changelog: v0.10.0...v0.11.0