The aiappdex.com model index pulled in 63 new models this week. I track additions by diffing models.json against last week's snapshot — new entries sorted by download count, not recency, so the list reflects actual usage rather than announcement order.
Here are the five that stood out.
Qwen3.8-27B-GGUF (Unsloth) — 3.56M downloads
The download count is high because GGUF quantizations are the format people actually use for local inference — llama.cpp, LM Studio, Ollama, and Jan all consume GGUF natively. Unsloth's imatrix-calibrated variants consistently rank above vanilla conversions because calibrated quantization preserves more of the original model's accuracy. 3.56 million downloads this week puts it at the top of new additions by a significant margin.
The base model is Qwen3.8-27B, Alibaba's 27-billion-parameter variant from the Qwen 3.8 series. The .8 notation indicates an architectural update within the Qwen3 line. At 27B parameters in a quantized GGUF, this runs on machines with a 24GB GPU or split across CPU+GPU.
1,840 likes on HuggingFace suggests this is getting real adoption, not just automated crawling. The 2.4× likes-to-download ratio is lower than average, which tracks for a utility file download rather than a model people are bookmarking to revisit.
MiniMax-H3 — 2.85M downloads
This is the most technically interesting model in this week's list. MiniMax-H3 is an image-text-to-video pipeline that generates synchronized audio and video from a text or image prompt — the audio isn't a separate post-process step, it's generated together with the visual content. The model card notes support for text-to-video, image-to-video, and video-to-video transformation.
4,149 likes with 2.85M downloads gives a likes-to-download ratio of roughly 1:690 — unusually high engagement for a video model, where most users are in experimentation mode rather than production deployment. MiniMax (the company) has been releasing models via HuggingFace as ComfyUI-compatible checkpoints, which accounts for the high download count — ComfyUI nodes pull weights automatically on first run.
The synchronized audio generation is the part worth watching. Most video generation pipelines treat audio as an afterthought. If MiniMax-H3 produces usable ambient audio without a separate TTS or music generation step, it simplifies a class of content production pipelines considerably.
Krea-2 (ComfyUI) — 2.10M downloads
Krea is an image and video generation product that's been running a web UI since 2023. This HuggingFace entry is the weights packaged for ComfyUI, distributed through comfy-org's official account. 2.1M downloads with 455 likes suggests automated node installation rather than manual research downloads.
The pipeline_tag is null in the registry, which usually means the model doesn't fit neatly into HuggingFace's classification schema — common for newer multimodal models or ComfyUI-specific formats that mix generation and upscaling in the same checkpoint.
What's notable here is Krea releasing weights at all. Their web product is commercial and they've historically kept the models proprietary. The ComfyUI release suggests either the weights behind the web UI are older versions being open-released, or this is a separate open track running in parallel with the commercial product.
Qwen3-30B-A3B-Thinking-2507-GGUF (Unsloth) — 1.27M downloads
Another Unsloth GGUF quantization, this time of Qwen3-30B-A3B-Thinking — a 30-billion-parameter Mixture-of-Experts model with a built-in chain-of-thought reasoning mode. The A3B in the name means 3B active parameters at inference time (out of 30B total), which is the key MoE efficiency trade-off: you get 30B-class capability at roughly 3B inference cost per token.
The Thinking suffix indicates this variant has extended reasoning enabled by default — similar to o1-style step-by-step generation rather than direct answer output. For tasks where chain-of-thought helps (coding, math, structured extraction), this performs well above what the 3B active parameter count would suggest.
1.27M downloads and 145 likes. Lower likes than the Qwen3.8-27B entry — probably because the Thinking mode is more specialized and the use case is narrower than a general-purpose LLM.
VAE + Text Encoder for FLUX Klein 9B (ComfyUI) — 1.20M downloads
This is a support component rather than a standalone model: the VAE and text encoder weights needed to run FLUX Klein 9B inside ComfyUI. FLUX Klein is a smaller, faster variant of Black Forest Labs' FLUX architecture. ComfyUI users typically download VAE and encoder weights separately from the main UNet — this bundle packages what's needed for FLUX Klein in one place.
1.2M downloads with 185 likes. The use pattern here is "I want to run FLUX Klein locally and this is the missing piece" — high download, low curation engagement, which is the typical profile for infrastructure components.
What I'm tracking next week
The video generation space is moving fast enough that MiniMax-H3's synchronized audio output is worth watching more carefully. If the quality holds for longer clips, it closes a pipeline step that currently requires three separate models (video generation + TTS + audio-visual sync).
The Krea weight release is the other signal. If their web product starts releasing weights systematically, it shifts Krea from a commercial API into the same open-weights category as Stability AI and Black Forest Labs — which means the aiappdex index will carry it as a citable open alternative to other video generation services.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)