Opening
On November 17, 2026, our radar picked up Haru-4B, a 4B-parameter audio-to-audio model with 4.4k downloads under Apache-2.0. The official model card claims 16GB VRAM for BF16, 8GB for 4-bit GGUF, and the ability to turn English meeting recordings directly into Chinese speech without text as an intermediate step. Also coming up: Inkling-2-Small, a compact vision-language model with 18.4k downloads that reportedly fits in 3.2GB at 4-bit GGUF, worth a look for engineers running invoice parsing on consumer GPUs; and fluid-speech-pro, a 0.5B ASR model that the authors clocked at 0.25 real-time factor on an Intel i7-12700 CPU, even running offline transcription on a Raspberry Pi 5, under CC-BY-4.0 for unrestricted commercial use.
This Week's Picks
1. higurashi-ai/Haru-4B: 4B End-to-End Speech-to-Speech, Commercial-Friendly, Fine-Tunable
- What it is
A 4B-parameter audio-to-audio model for Speech-to-Speech tasks, licensed Apache-2.0, commercial use and fine-tuning both allowed. The model card includes a fine-tune script and data format docs, so you can fork and experiment with LoRA for specific speaker styles (e.g., customer-service tone) right away. - Where it fits
- Indie developers wanting an offline "speech-to-speech" demo on a laptop (16–24GB VRAM), e.g., turning English meeting recordings directly into Chinese speech output, no text intermediary.
- Voice assistant product teams deploying low-latency multilingual conversational entry points on edge devices (Jetson Orin, Mac Studio); the model card notes streaming inference support.
- Grad students fine-tuning specific speaker styles with LoRA; the model card provides the fine-tune script and data format, ready to fork.
- Can it run?
Official model card states: "BF16 on 16GB VRAM, 4-bit GGUF on 8GB VRAM"; also provides GGUF quantizations (Q4_K_M, Q8_0) for llama.cpp. If you only have 8GB VRAM, the official 4-bit GGUF path is likely the easiest entry point. - Buzz this week
HF trending score 374, 4.4k downloads, 476 likes. - vs. peers
The author compares against SeamlessM4T-v2-large (1.2B) and Whisper + TTS pipelines in the model card, claiming Haru-4B wins on "end-to-end latency" and "tone preservation"; this is self-reported, no independent third-party reproduction yet. - How to get it
README gives two paths: ①pip install -U transformers acceleratethenAutoModelForSpeechSeq2Seq.from_pretrained("higurashi-ai/Haru-4B"); ② download GGUF withllama.cpp --model haru-4b-q4_k_m.gguf --audio-in mic. - Link
https://huggingface.co/higurashi-ai/Haru-4B
2. thinkingmachines/Inkling-2-Small: Compact VLM with 18.4k Downloads, 3.2GB at 4-bit
- What it is
A small vision-language model, Apache-2.0 licensed, supports 1024×1024 resolution, accepts PDF-rendered images directly. The model card includestrain_lora.pyandchat_templateexamples, ready to plug into Unsloth workflows for fine-tuning on your own corpus (e.g., traditional Chinese financial reports). - Where it fits
- SaaS backend engineers running "invoice/receipt auto-parsing" on 12GB VRAM machines; model card notes 1024×1024 support and direct PDF-rendered image input.
- Edge device engineers quantizing to 4-bit GGUF for deployment on Jetson Orin NX (16GB) for real-time "screen-content QA" test assistance.
- Data scientists LoRA-fine-tuning on proprietary docs (traditional Chinese financial reports); model card provides
train_lora.pyandchat_templateexamples, directly compatible with Unsloth.
- Can it run?
Official model card: "FP16 needs 8GB VRAM, 8-bit needs 5GB, 4-bit GGUF ~3.2GB memory"; pre-quantized GGUF (Q4_K_M, Q5_K_M) provided. This means 8-bit should theoretically run on 8GB consumer GPUs, and laptops with enough system RAM could even run 4-bit GGUF on CPU. - Buzz this week
HF trending score 347, 18.4k downloads, 434 likes. - vs. peers
Author compares against Qwen2-VL-2B and MiniCPM-V-2_6 in the model card, claiming 3–5 point gains on DocVQA; self-reported numbers, no public third-party reproduction yet. - How to get it
README instructs:from transformers import AutoModelForImageTextToText; model = AutoModelForImageTextToText.from_pretrained("thinkingmachines/Inkling-2-Small", load_in_8bit=True, device_map="auto"); or load GGUF withllama.cpp. - Link
https://huggingface.co/thinkingmachines/Inkling-2-Small
3. FluidVoice-Labs/fluid-speech-pro: 0.5B ASR Built for CPU Inference, Runs Offline Transcription on Pi 5
- What it is
A 0.5B-class model specialized for Automatic Speech Recognition (ASR), CC-BY-4.0 licensed, commercial use and redistribution allowed, no application needed. Model card provides ONNX and CTranslate2 versions, emphasizing "CPU real-time factor < 0.3". - Where it fits
- Full-stack engineers running "auto-generate podcast transcripts" in GPU-less CI/CD pipelines; model card provides ONNX and CTranslate2 versions, ready to
pip install ctranslate2for inference. - Embedded engineers deploying offline mixed Chinese-English recognition on Raspberry Pi 5 / Orange Pi 5 Plus; model card cites
int8quantization to 40MB, CPU real-time factor 0.25. - Privacy-sensitive teams (medical/legal) batch-transcribing recordings on air-gapped servers; CC-BY-4.0 permits commercial use and redistribution without application.
- Full-stack engineers running "auto-generate podcast transcripts" in GPU-less CI/CD pipelines; model card provides ONNX and CTranslate2 versions, ready to
- Can it run?
Official model card: "FP16 0.5B ~1GB VRAM; INT8 quantized CPU real-time factor 0.25 (Intel i7-12700), 0.35 (Apple M2); ONNX Runtime / CTranslate2 both supported". Machines without discrete GPUs, even Pi 5-class devices, can theoretically run offline as long as CPU performance suffices and memory fits the 40MB model. - Buzz this week
HF trending score 49, 9.1k downloads, 53 likes. - vs. peers
Author compares against Whisper-large-v3 (1.5B) and distil-whisper-large-v3 (756M) in the model card, claiming WER on CommonVoice zh-TW test set within 0.8% while being 4x faster; self-reported, no public third-party reproduction yet. - How to get it
README lists three ways: ①pip install fluid-voicethentranscribe("audio.wav"); ② download ONNX withonnxruntime-genai; ③ download CTranslate2 model withctranslate2.Translator("model_ct2"). - Link
https://huggingface.co/FluidVoice-Labs/fluid-speech-pro
Closing
All three models this week sit at the intersection of "small parameter count, solid quantization, permissive licensing": Haru-4B brings speech-to-speech down to a single consumer GPU, Inkling-2-Small compresses document understanding to 3.2GB, and fluid-speech-pro pushes transcription onto GPU-less CI/CD and Pi 5. If you have 16GB+ VRAM and want to try real-time speech translation, grab Haru-4B's GGUF for a quick feel; if you have zero GPU and just need transcripts, fluid-speech-pro's Python package is the lowest-friction path. See you next time.


























Comments