charls-nemotron-streaming-3.5-q8

Charls model provenance and mirroring

Charls ships exactly two transcription models. Both are mirrored from their upstream sources into Charls-controlled Hugging Face repositories before being referenced by the app's embedded catalog. This document is the audit trail: it records where each artifact comes from, under what license, at which pinned commit, and with what verified file hash.

The machine-readable source of truth is source-manifest.json, checked by bun run check:models (scripts/check-model-manifest.ts).

Models

1. Nemotron Streaming 3.5 Q8 (default, recommended)

• Base model: NVIDIA Nemotron Streaming ASR 3.5 (0.6B streaming ASR model), licensed under the Open Model Development and Weight License 1.1.
• Conversion/mirror repo consumed by Charls: handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf (GGUF conversion published by the upstream Handy project).
• Pinned source revision: 6d44e540bc31b0de1dbe174a3cea87f53a7f22fb
• File: nemotron-3.5-asr-streaming-0.6b-Q8_0.gguf
• Size: 751,094,240 bytes
• SHA-256: b94545b313b3223fda7b2857a52681da813935c2127643d1e9ff0c23d988089c
• License: Open Model Development and Weight License 1.1

2. Whisper Small F16 (fallback, lower-resource)

• Base model: OpenAI Whisper small, licensed under MIT.
• Conversion/mirror repo consumed by Charls: handy-computer/whisper-small-gguf (GGUF conversion published by the upstream Handy project).
• Pinned source revision: c0214bd34be9296695486f838e0142f900803159
• File: whisper-small-F16.gguf
• Size: 492,888,480 bytes
• SHA-256: bef65e1ac9d012269453243243aac0d0f67792693ba6c99b584a124e0ee326fc
• License: MIT

Publishing gate

Charls never serves models from handy-computer/* or blob.handy.computer in a release build. Instead, each artifact above is copied byte-for-byte into a Charls-controlled Hugging Face organization (${Organization}/charls-<slug>) via scripts/publish-charls-models.ps1 (see below), and the app's catalog is generated only from the resulting models/charls-lock.json.

Publishing an artifact to a Charls-controlled repository is forbidden unless the required upstream license files (LICENSE, attribution/origin notice) are present in the mirrored repository alongside the model file. The publish script uploads LICENSE, README.md, and ORIGIN.md together with the model artifact for exactly this reason — a repository missing any of these must not be treated as a valid publish target, and the operator must not point the catalog generator at it.

Operator procedure (scripts/publish-charls-models.ps1)

The publish script is the only supported way to produce models/charls-lock.json. It never touches handy-computer/* for anything other than reading the pinned source artifact, and it never uploads anything until size and SHA-256 have both been verified locally.

1. Dry run (no credentials, no network, no side effects)

Anyone can run this — it only validates the manifest, the lock schema, and the script's own inputs:

# pwsh (PowerShell 7), if installed:
pwsh scripts/publish-charls-models.ps1 -Organization charls-test -VerifyOnly

# Windows PowerShell 5.1, if pwsh is unavailable:
powershell -File scripts/publish-charls-models.ps1 -Organization charls-test -VerifyOnly

Expected output: 2 model definitions verified, exit code 0. Nothing is downloaded, nothing is uploaded, and models/charls-lock.json is not written or modified.

2. Real publish (owner-only, requires credentials)

Real mirroring is a one-way, owner-gated action deferred to whoever controls the Charls Hugging Face organization. It requires:

• The hf CLI (huggingface_hub) installed and on PATH.
• hf auth login already run with the Charls owner account (the script calls hf auth whoami as a strict preflight and aborts if it fails).
• A Hugging Face organization (or namespace) the owner controls, passed as -Organization. If the desired name is taken, stop and pick the final organization before running the script or editing the catalog — destination repo ids are not meant to be renamed after publish.

pwsh scripts/publish-charls-models.ps1 -Organization <CharlsHuggingFaceOrganization>

For each of the two models in source-manifest.json the script will:

1. Download the pinned file from the source repo at the pinned source_revision.
2. Verify its size and SHA-256 against the manifest and abort before any upload on a mismatch.
3. Download the upstream LICENSE file from the source repo and refuse to publish if it is missing.
4. Create (or reuse) ${Organization}/charls-<slug> and upload the model file plus LICENSE, README.md, and ORIGIN.md.
5. Query https://huggingface.co/api/models/<repo_id> for the resulting commit sha and record it as the pinned destination revision.

The resulting lock is written atomically (models/charls-lock.json.tmp then renamed to models/charls-lock.json) and must validate against charls-lock.schema.json: exactly two entries, each with a repo_id, a 40-character revision, a filename, an integer size_bytes, and a 64-character sha256.

models/charls-lock.json is the only input the catalog generator (scripts/gen_catalog.py, Plan 3 Task 3) is allowed to read destination repo ids and revisions from — it must never read source-manifest.json's source_repo/source_revision directly.