Glimpse C1 2B
C1 is the cleanup model in Glimpse. It turns raw dictation into clean text, entirely on your computer.
- Filler and false starts. Ums and abandoned restarts disappear.
- Self-corrections. "At two, no actually three" becomes "At three."
- Punctuation. Sentences come back with the marks in the right places.
- Spoken commands. "Make this a bullet point list" makes a list.
- Modes. Follows writing instructions, from a formal email to a quick chat message.
C1 is a LoRA adapter on Ternary Bonsai 2 2B from PrismML, a ternary model derived from Qwen3.5. Read the announcement.
Results
Glimpse's cleanup tests, run on an M2 Pro with each model's recommended prompt.
| Model | Cleanup suite (171) | Self-corrections (24) | Raw speech (52) | Non-English (55) |
|---|---|---|---|---|
| Glimpse C1 2B | 162 | 22 | 47 | 54 |
| Superwhisper S1-mini | 123 | 20 | 39 | 35 |
| SimpleWords v2 | 118 | 15 | 24 | 40 |
| Apple Intelligence | 114 | 15 | 27 | 36 |
| Bonsai 2B, untrained | 88 | 1 | 15 | 33 |
A typical cleanup takes 0.27 s on an M2 Pro.
Files
| File | |
|---|---|
adapter.safetensors |
The C1 adapter (alpha / r folded into B) |
model.safetensors |
PrismML's MLX 2-bit Bonsai 2B, vision tensors removed |
config.json, tokenizer.json |
From PrismML's pack |
LICENSE, NOTICE.txt |
PrismML's Apache 2.0 license and notice for the base model |
LICENSE-C1.md |
The license for the adapter |
Usage
C1 runs in Glimpse on Mac (Metal) and Windows (Vulkan). It was trained on Glimpse's cleanup prompt, with the dictation in the user turn as <transcript>…</transcript>. With other prompts, expect weaker results.
Limitations
- It is a 2B model. Long or heavily tangled dictation can come back with a correction missed or a phrase kept that should go.
- Modes are followed on about 80% of held-out cases.
- It is tuned for cleanup. It is not a chat model.
License
The adapter is free to use through the Glimpse app, including for commercial work, and free for personal use on your own devices. Shipping it in another product or service, sharing or hosting it, other commercial use, or training on its outputs needs written permission from Glimpse: hello@tryglimpse.cc. See LICENSE-C1.md.
The base model files are PrismML's, under Apache 2.0 (see LICENSE and NOTICE.txt). Created using Bonsai by Prism ML.
- Downloads last month
- 1,466
Model tree for Glimpse-Dictation/Glimpse-C1-2B
Base model
Qwen/Qwen3.5-2B-Base