---
license: mit
tags:
  - not-a-model
---

# SpeakEasy Mini — CUDA runtime worker (Windows x64)

**This repository contains no model.** It carries one compiled Windows
executable: the CUDA-enabled build of SpeakEasy Mini's Granite Speech inference
worker. The model it runs is
[ibm-granite/granite-speech-4.1-2b-GGUF](https://huggingface.co/ibm-granite/granite-speech-4.1-2b-GGUF),
which lives elsewhere and is fetched separately.

SpeakEasy Mini is a local-only Windows dictation app. All inference runs on the
machine; nothing is sent anywhere. Its installer fetches this archive only when
the user chooses the graphics-card configuration on a machine with a supported
NVIDIA card, and verifies it against the digests below before it is used.

## What is in here

| | |
| --- | --- |
| `granite-worker-cuda-windows-x64.zip` | 42,162,465 bytes, SHA-256 `03b7c33f83674edfa64549e43a5f87e3393e03100db44027b10940b5deb28147` |
| contents | `granite-worker.exe`, 57,052,672 bytes, SHA-256 `1d4a3ad57e72acaaa55a507f88733ac18f77909349c4bf7c7bc017269499170a` |

The archive has no wrapping directory. The digests above are pinned in the
application's own trusted manifest and re-checked on the user's machine after
download and again before the worker is run, so a corrupted or truncated
transfer is refused rather than executed.

## Runtime requirements

The executable is a **CUDA build** and does not fall back to the processor. It
needs, beside it in its own directory:

- `cublas64_13.dll` and `cublasLt64_13.dll` — from NVIDIA's CUDA 13.x
  redistributables. `cuBLAS` is an import of the image; `cuBLASLt` is loaded by
  `cuBLAS` at the first matrix multiplication, so a copy missing only that one
  starts, loads the weights, and fails roughly half a minute later.
- an NVIDIA driver providing `nvcuda.dll`.

The CUDA runtime itself is statically linked, so `cudart64_13.dll` is not
loaded. SpeakEasy Mini's installer fetches the two libraries above from NVIDIA's
own distribution servers rather than from here.

## Build provenance

- Source: [`kwp490/speakeasy-granite-rust-mini`](https://github.com/kwp490/speakeasy-granite-rust-mini)
  at commit `67c94987d26983767ef1ceb05f7f62fdbd902858`
- Built with `cargo build --release -p speakeasy-granite-worker --features cuda`
  on Windows, MSVC toolchain, against CUDA Toolkit 13.3
- Verified on an NVIDIA RTX 4070 Laptop GPU: transcribed the project's fixture
  clip in 361.2 ms resident, byte-identical to the processor build's transcript,
  with NVML confirming the worker's own process holding a CUDA context

MSVC embeds a timestamp and a PDB path, so this build is not bit-reproducible. A
rebuild from the same commit will produce a different SHA-256. The digests above
identify these bytes, which are the ones that were tested.

## Licences

See [`NOTICE.md`](NOTICE.md) for the full notices. In summary, this executable
contains:

- **llama.cpp / ggml** — MIT, Copyright (c) 2023-2026 The ggml authors
- **nlohmann/json** and **cpp-httplib** — MIT, as vendored by llama.cpp
- **llama-cpp-rs** bindings — MIT or Apache-2.0, Copyright (c) Dial AI
- **NVIDIA CUDA runtime**, statically linked — governed by the
  [NVIDIA CUDA Toolkit EULA](https://docs.nvidia.com/cuda/eula/index.html)
- SpeakEasy Mini's own code — see the source repository

The `mit` tag in this page's metadata describes the project's own code and its
MIT-licensed dependencies. It does not extend to the statically linked NVIDIA
components, which remain subject to NVIDIA's own terms.