Opening
In the November 16, 2026 scan of the local model radar, we dug up a serious contender: tiny-voca, weighing in at just 18.2MB, with the team claiming it can run at 70x real-time on CPU alone, no GPU required. The other headliner is FLUX.1-schnell, a 12B-parameter model released under the Apache-2.0 license for open commercial use and LoRA training. Let's dig into how to pick between these two.
This Episode's Picks
1. tiny-voca: A Speech Model Small Enough to Bundle Right into Your Repo
-
What it is: n4ze3m/tiny-voca is a text-to-speech (TTS) model with only 4.5M parameters and an extremely small weight file, released under the Apache-2.0 license with no restrictions on personal or commercial use. Compared to typical TTS models that start at hundreds of megabytes, this one is a genuine outlier in the lightweight category.
-
Where it fits:
- If you're building a small browser-based tool or something for an edge/IoT device and want to add voice feedback without complicating your deployment with a GPU server, this model is small enough to bundle directly into your repo and ship as-is, and the Apache-2.0 license means you don't have to worry about hitting commercial-use roadblocks.
- If you're adding an English voice alert or announcement feature to a side project, say, a timer that says "time's up," or a small robot that talks back, you can solve it with a model under 20MB that doesn't need to run in the cloud, without renting extra GPU capacity for one small feature.
- When you need a lightweight baseline for comparing TTS model quality, this one comes with publicly verifiable WER rankings, which makes it a more credible reference point than just grabbing a random small model off the shelf.
-
Will it actually run: The official model card states "runs about 70x faster than real time on four CPU threads," with no GPU required. The model file is only 18.2MB (PyTorch format) or 21.4MB (ONNX format), and GGUF is also available as a format option. Whether you're running PyTorch, ONNX Runtime, or another lightweight inference environment, there's a format that fits.
-
Current buzz: HF popularity score of 135, 860 downloads, 143 likes. The numbers aren't jaw-dropping, but for a recently released 4.5M-parameter model, there's already a solid group of people using it.
-
How it stacks up: According to the Open TTS Leaderboard (November 2026), tiny-voca's M2 voice "currently ranks 8th by average WER," and its F2 voice "ranks 13th." It's rare to find a model at this 4.5M-parameter scale with verifiable third-party benchmark rankings, rather than just the model's own self-reported claims.
-
How to get it: The official docs provide an install command,
pip install git+https://github.com/n4ze3m/dizione git+https://github.com/n4ze3m/tiny-voca, after which runningpython -m tiny_voca.infer --checkpoint tiny-voca/tiny-voca.pt --voice F2 --text "Hello"generates speech. An ONNX version is also available, loadable via onnxruntime.
2. FLUX.1-schnell: An Image-Generation and LoRA Training Base Without Licensing Headaches
-
What it is: black-forest-labs/FLUX.1-schnell is a 12B-parameter rectified flow transformer for image generation, released under the Apache-2.0 license. According to the official docs, it was trained using latent adversarial diffusion distillation, allowing it to generate images in just 1 to 4 inference steps, significantly faster than traditional diffusion models that typically require 20 to 50 steps.
-
Where it fits:
- If you want to set up an image-generation service on your own machine, whether through local ComfyUI or your own custom pipeline, and don't want to get tripped up by licensing terms, the Apache-2.0 schnell version places no restrictions on personal or commercial use, so you can adopt it directly as your base model.
- If you want to train your own style LoRA, say, for brand visuals or a specific illustration style, FLUX.1-schnell is currently one of the most commonly used free base models in the community for LoRA training, making it relatively easy to find existing tutorials and community resources.
- When you need to prototype images quickly, for example, generating a few concept images for colleagues before a design review, and don't want to wait through a full round of traditional diffusion steps, the official example only requires setting
num_inference_steps=4to produce output, cutting wait time considerably.
-
Will it actually run: The official model card doesn't specify a clear VRAM threshold, but the Diffusers official example provides an option,
enable_model_cpu_offload(), that lets users "save VRAM by offloading the model to CPU," suggesting the team also anticipated this model being used in resource-constrained environments. However, the exact amount of VRAM needed for smooth operation isn't explicitly stated. Safetensors is the available format. -
Current buzz: HF popularity score of 55, 280.1k downloads, 6.8k likes. Among this batch of picks, the download count here is particularly solid.
-
How it stacks up: The official docs state the model was trained using latent adversarial diffusion distillation, enabling image generation in 1 to 4 inference steps, and the model card describes its performance as reaching "the performance of closed-source alternatives." However, it doesn't provide direct benchmark comparisons against FLUX.1-dev or SDXL, so this part is based solely on the official description, without a quantitative comparison table.
-
How to get it: The official example loads it via Diffusers:
from diffusers import FluxPipelinecombined withFluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16), and calling it withnum_inference_steps=4, guidance_scale=0.0generates an image.
Closing
Today's two picks sit at opposite ends of the local model ecosystem: tiny-voca proves that speech models don't need a GPU to work, with its minuscule footprint, while FLUX.1-schnell uses a permissive license to clear the path for image generation and LoRA training. If all you've got is a laptop without a GPU, give tiny-voca a try first. If you've got a graphics card and want to train your own style LoRA, FLUX.1-schnell is worth putting on your shortlist. Same time tomorrow, we'll keep digging through the pile of local models. See you in the next episode.


























Comments