Repository navigation
Vella 2.0.0
-
Tensor kernels require macOS 26.2 or newer. Optimized-path verdicts now include GPU architecture and device name; existing verdicts requalify once when each model/recipe next loads.
-
vella statusnames each loaded model's tier and effective Standard/Optimized Fast/Optimized Exact path; stock fallbacks retain the requested recipe in parentheses. -
The Microphone menu checks the system default input at recording start and can use an available fallback. Losing the mic mid-recording stops capture, keeps the audio and offers Retry.
-
The bundled schema-2 diagnose reference is the final quiet-window qualification, with recorded conditions. Release checking and packaging validate it against the shipped model recipes.
-
Known limitation:
--languagedoes not change recognition in 2.0; it only setsverbose_json.language, which echoes the requested code, orunknownwhen omitted. Detected-language reporting is deferred to 2.1. -
Kept performance levers are now on by default for their measured model/precision cells: native INT8/INT4 in Parakeet Fast; tail blocks in Ultra Fast and Exact; keep-cache in Nemotron Fast and Exact, plus joint batching in BF16 Fast. Standard uses none. Lab
=1/=0overrides remain; effective lever configurations match measured verdicts; GPU-identity keys require fresh local qualification. -
Installers refuse older versions and older builds of the same version unless explicitly authorized with
--allow-downgrade. Signing checks still apply; manual backup restoration is unchanged.
Known limitations
- Streaming memory: in 20-minute real-time tests Nemotron's peak footprint was 1.78 GB at 16-bit and 1.19 GB at 8-bit, ending below its loaded level. These tests do not bound long-duration memory growth. Streams of an hour or more are not yet verified.
Added
- Prebuilt install.
scripts/install.shdownloads the release zip with curl, verifies its SHA-256, contents, version and code signature before touching anything, swaps the app in place and waits until it is ready (ready: …).scripts/install-release.sh <version> --dry-runverifies without installing. No Xcode or Python needed;VELLA_BUILD=sourcebuilds the checkout instead. - Keep Hot per kind of load: manually loaded models stay loaded (default Always) and load again at launch; models loaded for a dictation unload after 15 idle minutes by default.
- Memory: Fit in free memory (default) unloads idle models or refuses a load with the numbers and remedies; Allow swap (slower) skips the check.
- Engine label
Optimized Fast · <chip>,Optimized Exact · <chip>orStandardunder a loaded model, backed by a self-test at load and a fallback that redoes a failed optimized transcription on the stock path. - First dictation without a model keeps the recording and offers Get (); the recording is transcribed once the model is ready.
- In-app updates. A newer release shows an orange Update to X… item under Support the Developer…. Update Now downloads the release, verifies its SHA-256, contents, version and that it is signed like the running app, installs it when Vella is idle and restarts; if the new version does not become ready, the previous one is restored. It replaces the notice that only opened the release page.
- Audio files, for you and your agents.
vella transcribe <file>(text,--srt,--vtt,--json) and an OpenAI-compatible local API (POST /v1/audio/transcriptions,GET /v1/models,GET /status) on 127.0.0.1, so the OpenAI SDKs work with only the base URL changed. Any audio macOS decodes, up to 3 hours; dictation always goes first; nothing is pasted or kept. The installer linksvellainto~/.local/bin; Copy Skill for Your Agent andvella skillprovide the agent skill. - One-command install without git.
curl -fsSL https://tobynoskillson.github.io/Vella/install.sh | bashnow installs the prebuilt release with the same checks asscripts/install.sh(it built 0.8.8 from source with Python).bash -s -- --dry-runverifies without installing. - Benchmark site at https://tobynoskillson.github.io/Vella/: every measured model and precision on the v2 benchmark, sortable, with the estimated cloud API rows.
- Models controls for agents.
vella models --jsonlists all table rows and cells; Select, Get (explicit--yesconsent), Load, Reload, Unload, Delete (explicit precision and--yes), Keep Hot and Memory use the table's controller and runtime. Authenticated local API routes expose the same controls; file transcription itself still never downloads. - Optional disk image.
scripts/package-dmg.shmakes a localVella-2.0.0.dmgfrom the verified release ZIP, keeps the app's existing signature and adds its checksum. The primary installer still uses the ZIP.
Measured
-
Every measured precision is offered. All 63 model × precision × path cells can be selected, with their measured figures. 21 of them fail the presence check against 16 bits: test clips came back empty or cut short (Qwen3 ASR 1.7B int8 also loses 42.64 points in Turkish). Their tooltips and
vella models --jsonsay what failed, and they are never recommended. -
Nemotron gates are layout-aware. Each gated cell compares with native Standard using the same session/shard order and endpoint gap; a missing same-layout control is not a pass. Int8 and int4 are offered but not recommended; int4 loses 21 test clips.
-
On Apple M5 Max, macOS 26.6, 1–2 October 2026: Parakeet v3 Ultra bf16 · Optimized Fast is 15.51% English WER, 507.1× real time, 4.58 J per audio minute and 1,792 MB peak RAM. Accuracy uses v2 (239.7 min); speed, energy and peak RAM use v2-quick (22.5 min). Measured build deb8845 was built from 6eed904; the final figures carry the source bridge and build provenance in
Resources/benchmarks.json. -
Whisper Standard and Optimized were measured in the same quiet window on the faithful FP16 build; the refresh completed 4 October. Exact and Fast use the same exact-only whisper-4 recipe and canonical measurement. Per-cell and tier gates use that build’s FP16 Standard baseline; the earlier Float32-baseline figures remain withdrawn.
Changed
- A fresh install downloads and loads nothing. Earlier versions downloaded and selected Parakeet Q4 during installation.
- One helper process per loaded model; a crashed helper restarts up to three times (after 2, 4 and 6 s). Helpers exit when Vella quits or is force-quit, and helpers left behind by an earlier Vella are stopped at launch, matched by their executable file only.
- The installer keeps the previous app until the new one reports ready, and refuses while a model is loading as well as during dictation.
- Signing migration. An installed Vella with an ad-hoc, Apple Development or other verified signature can move to the pinned release identity only with explicit
--migrate-signingconsent. The installer names the identity being replaced, asks y/N only with terminal stdin, keeps the old app for rollback and warns that macOS will ask for Microphone and Accessibility again. Same-identity updates need no migration consent. For 0.8.x or development-signed installs:curl -fsSL https://tobynoskillson.github.io/Vella/install.sh | bash -s -- --migrate-signing. - Updates replace the whole app bundle, so files from older versions (including Python-era resources) never survive an update.
- Licence changed to MIT. Vella 2.0 is licensed under the MIT License. Published 0.x releases remain Apache-2.0; their licences are not changed retroactively. Third-party code and model weights retain their own licences.
- Chip coverage. Standard is optimized for your Mac through MLX; Optimized adds our custom kernels. Measurements are from an M5 Max with a 40-core GPU; other M5 configurations have not yet been tested. The CLI help and agent guides state that boundary explicitly.
- Installation. The release installer remains the primary route: checksum and signature verification, quarantine removal and automatic launch. The standalone installer now checks macOS 26, matching the shipped app.
scripts/install.sh --dry-runnow forwards verification-only mode correctly. - Updating legacy quantizations. Earlier published 8/4-bit downloads are kept on disk but no longer used as measured tiers. When the same tier is offered and its 16-bit source is installed, a saved selection moves to the measured local derive at that tier, without a download. A tier no longer offered moves to the installed 16-bit source. Every changed selection is disclosed in a visible launch notice and one log line per mode. Without that source, the selection is cleared; its reason is retained and shown when the next recording needs a model. Use Get for its 16-bit source. Retained legacy files, including tiers no longer offered, can be deleted only after confirmation in Models or with
vella delete ID --precision int8/int4 --yes. - The dictation helper keeps up to 64 MB of reusable GPU buffers between requests instead of freeing them after every request, so dictation is slightly faster (about 4 % on Parakeet v3 Ultra BF16). It never holds more than 64 MB while idle, frees them after a failed request, and releases them on unload or when macOS reports memory pressure.
VELLA_DICTATION_KEEP_CACHE=0restores the old behaviour. - Smaller app and download. Dictation and live-transcription helpers are now one program, and shipped programs carry no debug symbols. Both helper processes appear as VellaWorker in Activity Monitor,
topand crash reports (signing identifierVellaWorker); the streaming one is still started asVellaStreamingWorker. Use a PID rather than a process name for diagnostics. Release packages include a separate, checksummedVella-VERSION-arm64-symbols.zipwith UUID-matched dSYMs for crash symbolication; installers download only the app. Plain source builds retain no extra symbol directories. - Only catalog models load. Vella loads the models in its catalog (downloaded or converted through the Models table); a model folder chosen some other way is no longer accepted. At launch, a dictation or streaming selection that points outside the catalog is cleared (and noted in the log).
- Models table. Every model shows all six Precision cells, named by the format that runs (
bf16/fp16,int8,int4), on an Optimized row (a bolt) above a Standard row (the MLX logo); a cell the model cannot run is greyed in place with the reason in its tooltip. The Fast/Exact switch is as tall as both rows and sets the Optimized row. Memory is now Peak RAM. The Capabilities column is gone (each model's languages are in its name's tooltip), and a thick line divides the Dictation and Streaming groups. The Models menu takes the table's exact width, so the row buttons are no longer cut off at its right edge. The final release figures are shown, including same-window Whisper Standard and shared Exact/Fast measurements. Presence is cell-level: an unavailable path is greyed in place without disabling another offered path at that precision. - The optimized path's load self-test checks only the kernel class Vella uses, so a load qualifies sooner (about 0.6 s instead of 0.7 s on an M5 Max). Keys, verdicts and engine labels are unchanged.
- Removed development switches:
VELLA_UPDATE_CA_CERT,VELLA_QWEN_ENC_BF16,VELLA_QWEN_PREFILL_HEAD,VELLA_QWEN_HOST_LENGTHS,VELLA_QWEN_REFERENCE_LENGTHS,VELLA_WHISPER_ENC_F16,VELLA_WHISPER_FUSED,VELLA_PARAKEET_FP32_FRONTEND,VELLA_NEMO_GEMV_R,VELLA_NEMO_GEMV_S,VELLA_RENDER_SWITCH_WORDS. Parakeet loads TDT models only.VellaModelToolis a stub that prints a notice (kept for 1.0.x updaters). - Code reorganised by feature, with one model interface in the recognition helpers, a shared wire package (
Packages/VellaWire), Swift 6 language mode for the MLX-free libraries and tools, and SwiftLint and swift-format checks (scripts/lint.sh).
Fixed
-
Pipefail-safe scripts. Signing-identity, existing-signature and runtime-symbol probes drain producer output and refuse inspection errors rather than failing on a successful early match.
-
API file decode overlap. File audio decoding overlaps transcription of the preceding segment, while retaining dictation priority and the existing admission limits.
-
Whisper's Standard path is activation-dtype-faithful to mlx-whisper on shipped FP16 checkpoints. The FP16 checkpoints omit the encoder's positional table; Vella synthesised it in Float32, which made the whole encoder and decoder compute in Float32 and convert every FP16 weight on every call. Running in FP16 had been counted as an Optimized component; Optimized · Exact and Fast are now the same exact decoder components, and Optimized · Fast retains the old tensors on catalog checkpoints. Vella preserves its sinusoid values (about 1.85% differ from mlx-whisper by one FP16 ulp); bitwise reference equivalence is not claimed. Mel remains FP16-rounded for arbitrary source dtypes.
-
Short or silent endings no longer become separate requests. A dictation or file whose last segment had under 2 s of new audio, or no audio above the silence level, sent that piece to the model on its own: Whisper answered such pieces with "Thank you." and similar phrases, and Parakeet dropped the opening sentences of a clip cut 0.1 s before its end. That piece is now transcribed together with the segment before it. Saved recordings are written and cut exactly as before.
-
Whisper models get longer segments. Whisper was trained on 30 s windows; it now cuts at a pause only after 20 s (other models: 5 s), so it sees whole sentences and makes fewer errors at cuts.
Removed
- The Python source-build installer and runtime files. Benchmark corpora, raw results and developer test harnesses are no longer part of the repository.
Install or update with scripts/install.sh from a checkout of this tag (see the README): it downloads this zip and checks it against SHA256SUMS before installing. To check the zip's build provenance:
gh attestation verify Vella-2.0.0-arm64.zip --repo TobyNoSkillSon/Vella
Signed with Vella's own certificate "Vella Release Signing" (SHA-1 2CA2587C8B85EF687E68950E405EC58CE31FC1C7), the same in every release. Microphone and Accessibility permissions carry over between releases signed with this pinned identity; the first migration from a development or other identity requires granting them again.