S1-mini-Q8_0-Q6_K-GGUF

S1-mini · GGUF

Two GGUF builds of S1-mini by Superwhisper, the speech-transcript normalizer, made for Lathe, a local push-to-talk dictation app for Windows. Superwhisper publishes F16 and Q4_K_M at superwhisper/s1-mini-GGUF; these fill the gap between them.

Q4_K_M was observed dropping whole clauses from dictations; neither build here did on 202 test inputs. Where Q6_K mixed differs from F16 it is nearly always punctuation, or a filler removed that F16 kept.

Q8_0: llama-quantize on the upstream F16, no importance matrix (Q8_0 does not use one).

Q6_K mixed: Q6_K with an importance matrix calibrated on 266 real dictations in the model's own prompt format, then per tensor: the output head and the six blocks a sensitivity sweep found most fragile (0, 11, 16–19) at Q8_0; attn_v, attn_output, ffn_up, ffn_down at Q6_K; attn_q, attn_k, ffn_gate at Q5_K; the token embedding at Q4_K.

Full comparison of fourteen builds, including the recipes that did not beat plain Q8_0, in Lathe's docs.

Prompt format

S1-mini has a fixed contract: the system prompt below verbatim, then a control line and the raw transcript, thinking disabled, greedy decoding.

<|im_start|>system
You are a text normalizer for speech-to-text transcripts. The input begins with a control line specifying the styling, structure, and context settings; clean the transcript to match those settings and output only the cleaned text.<|im_end|>
<|im_start|>user
[Styling: semi-formal] [Structure: prose] [Context: general]
okay so can you send me the the file from yesterday not the older one thanks<|im_end|>
<|im_start|>assistant
<think>

</think>

Styling: casual, semi-casual, semi-formal, formal. Structure: prose, lists. Context: general, email.

License

Apache 2.0, inherited from S1-mini and from Qwen3-0.6B (Alibaba Cloud), with Superwhisper's additional term: any use or distribution of this model must continue to identify it as "S1-mini" by "Superwhisper". See LICENSE and NOTICE, copied unchanged from the upstream repo.