cloudkites/nbq-models

CloudKites nbq models

Open-weight models in .nbq format, the weight format CloudKites' on-device apps load with the numbat runtime. One copy here, shared by every app that runs the model.

file	base model	quantisation	size	SHA-256
qwen3.5-4b/qwen3.5-4b-q4km.nbq	Qwen/Qwen3.5-4B	Q4_K_M (by Unsloth)	2.75 GB	d5c3411cd851c6adcbd670ee5c0f9cbe37e4cf7b40deedb549259615e7834fb1
qwen3-asr-0.6b/qwen3-asr-0.6b.nbq	Qwen/Qwen3-ASR-0.6B	none (original precision)	1.88 GB	d42bc187c41fbcecd3cce62b4867fea2fb2a206cd1923642cb932772635c3d8a
omi-med-stt-v1-int8/omi-med-stt-v1-int8.nbq	omi-health/omi-med-stt-v1 (Parakeet TDT 0.6B v2 fine-tuned on medical speech)	int8 weights, per output row	668 MB	344fda6029ee39c7ab699a09398014947031d0597a61d18d79a5cb210a829d90

Licence

Each model keeps its own licence; all credit for the models belongs to their authors.

• Qwen3.5 4B and Qwen3-ASR 0.6B are Qwen's, released under the Apache License 2.0 (LICENSE in this repository, as published with Qwen/Qwen3.5-4B). Changes from the original: Qwen3.5 4B quantised to Q4_K_M by Unsloth; both repackaged as .nbq. The weights are otherwise unchanged.
• Omi Med STT v1 is by Omi Health, a fine-tune of nvidia/parakeet-tdt-0.6b-v2, licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Changes from the original: the weight matrices quantised to int8 (symmetric, one scale per output row), the rest kept in 32-bit float, and repackaged as .nbq. Omi's notice is in omi-med-stt-v1-int8/NOTICE.md. This is not an NVIDIA or Omi Health release, and neither endorses it.

Intended use

General-purpose language model weights for local, on-device inference. Not a medical device and not validated for clinical use; any application built on them carries its own responsibility for how they are used.