HF breakouts Sep 7–14: zero verified >10x jumps; four fresh watchlist signals
The weekly screen finds zero verified >10x Hugging Face breakouts and four fresh watchlist models whose licenses and deployment costs deserve a builder’s test.
Evidence cutoff: September 14, 2026 at 09:00 PT. The strict screen found zero verified Hugging Face models with more than 10x download growth during the Sep. 7–14 window.
The screen requires two dated, comparable
Downloads last month observations for the same repository. That metric is a rolling 30-day discovery signal, representing recent attention rather than an exact weekly cohort or unique-user count. The latest dated snapshots support four fresh watchlist signals across language, multimodal, and audio categories.| Model | Modality | Dated rolling-30-day snapshots | Strict result | License and commercial status | Builder fit |
|---|---|---|---|---|---|
Edge0/Edge0-35B-A3B-preview | LLM | 3,552 on Sep. 14; unlisted on Sep. 7 12 | No prior snapshot; watchlist | Apache 2.0. Commercial use is permitted under standard Apache 2.0 terms. 3 | A 35B MoE running in under 3 GiB RAM via SSD expert offload and predictive prerouting on Apple Silicon. 3 |
TokenRhythm/NeoHorse-1-4B | LLM | 7,979 on Sep. 14; unlisted on Sep. 7 12 | No prior snapshot; watchlist | Apache 2.0. Commercial use is permitted under Apache 2.0, with upstream Alibaba Cloud copyright preserved. 4 | A 4B agent-focused model post-trained from Qwen3.5-4B for tool use, coding, and lightweight local assistants. 4 |
deepseek-ai/DeepSeek-V4.1-Flash | Multimodal | 244,457 on Sep. 14; unlisted on Sep. 7 12 | No prior snapshot; watchlist | MIT License. Commercial use is permitted under standard MIT terms. 5 | A 552B multimodal MoE with 16B active decode parameters, 1M context, and 890 B/token KV cache efficiency. 5 |
m-a-p/YuE2-3B | Audio | 3,707 on Sep. 14; unlisted on Sep. 7 12 | No prior snapshot; watchlist | CC BY-NC 4.0. Model weights are restricted to non-commercial research and evaluation. 6 | Full-song stereo music generation with editable ABC scores and accompaniment control on a 24 GB GPU. 6 |
LLM
Edge0-35B-A3B-preview
Edge0-35B-A3B-preview addresses memory-constrained deployment by offloading sparse mixture-of-experts parameters to local storage. Built on Qwen3.6-35B-A3B with 256 experts and 4 active experts per token, the checkpoint stays on SSD storage while an MLX runtime streams active weights into RAM during execution. 3
The model pairs SSD expert offload with a trained prerouter head that predicts expert routing one step ahead. Benchmarks on an Apple M4 Pro Mac mini with 24 GB RAM report 2.9 GiB peak active memory and 14.9 to 17.7 tokens per second decode speed. Recover-LoRA distillation keeps the 4-bit checkpoint within 3.9 points of the fp16 base model across standard benchmarks. 3
The release carries an Apache 2.0 license, providing a clear legal basis for commercial deployments. The maintainers identify agentic workflows as an active limitation in this preview checkpoint: multi-step tool calls and planning remain weak. 3
The practical prototype test is a local document chat assistant on Apple Silicon hardware. Measure generation speed, SSD read throughput, and answer quality across 100 conversational prompts before evaluating the pipeline for production tasks.
NeoHorse-1-4B
NeoHorse-1-4B is a compact 4B causal language model tailored for agent execution. TokenRhythm post-trained the model from Qwen3.5-4B using routing-guided curriculum supervised fine-tuning and on-policy distillation. The checkpoint provides a native 262,144-token context window with extension capability up to 1,010,000 tokens. 4
Benchmark evaluations show strong tool-calling and code execution results, including 96.95 on HumanEval, 77.33 on PinchBench, and 88.46 on tau2-Bench. The authors report a 64.87 macro average across ten evaluations, representing a 5.93-point improvement over the base Qwen3.5-4B model. Runtimes including SGLang v0.5.17 and vLLM support the checkpoint through standard Qwen3 coder parsers. 4
The Apache 2.0 license covers commercial use, retaining upstream attribution notices for Alibaba Cloud. The weights support text-only inference; visual inputs are excluded from this release. 4
The recommended prototype test is a dedicated code-patching agent. Connect the model to an automated test harness with five failing unit tests, track tool invocation accuracy, and measure time to successful resolution on a single consumer GPU.
Image generation
The Sep. 7–14 window yielded zero verified text-to-image models meeting the paired-snapshot standard. The evidence set contains video generation activity such as Lightricks LTX-2.5, which recorded 1.548 million rolling downloads on Sep. 14 compared to 1.526 million on Sep. 7, reflecting stable usage rather than a breakout. 12
Builders seeking image-generation checkpoints should keep this category empty for the current cycle.
Audio
YuE2-3B
YuE2-3B is an open music generation model designed around editable musical representations. Built on an autoregressive-nonautoregressive Mixture-of-Transformers architecture, the system generates complete 48 kHz stereo songs with vocals and accompaniment from text prompts or user-supplied ABC scores. 6
The system operates locally on a single 24 GB GPU, completing a 3.6-minute stereo composition in 71 seconds on an RTX 4090 with 11.18 GiB peak VRAM. Evaluations on WildSongBench claim an average score of 6.9632, surpassing comparable commercial baselines. 6
Licensing defines the primary product boundary. The model weights are distributed under CC BY-NC 4.0, restricting usage to non-commercial research and private experimentation. Commercial product integration requires a separate commercial grant from the authors. 6
Builders can test YuE2-3B for internal audio ideation or creative prototyping. Supply a custom chord progression in ABC notation, record vocal alignment and accompaniment balance, and evaluate audio quality on an RTX 4090 workstation.
Multimodal
DeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash is a large multimodal Mixture-of-Experts model supporting interleaved text and visual inputs. The architecture uses a Causal Encoder-Decoder Transformer with a 552B parameter backbone, activating 8B parameters per token during prefill and 16B parameters during decode across 384 routed experts and one shared expert. 5
The model supports contexts up to 1,000,000 tokens while compressing global KV cache storage to 890 bytes per token. It introduces continuously controllable reasoning effort on a scale from 1 to 100, alongside strong multimodal benchmarks including 95.6 on DocVQA and 56.5 on MMMU-Pro. Managed inference endpoints on Novita, Fireworks, DeepInfra, and Baseten provide hosted access. 5
The repository and weights carry an MIT License, permitting commercial production deployments and self-hosting. 5
The ideal evaluation is a multi-page document intelligence task. Submit 20 financial statements containing dense data tables and charts, benchmark extraction precision against existing multimodal endpoints, and measure serving latency across varied reasoning effort settings.
Before building
- Test each candidate model on a representative domain task using an explicit pass-fail threshold.
- Measure peak RAM, memory bandwidth, latency, and hosting costs on the target deployment hardware.
- Verify repository-level licenses across model weights, tokenizers, runtime dependencies, and quantization scripts.
- Require two dated, independent Hugging Face download observations before treating any repository as a verified breakout.
References
- 1
- 2
- 3Edge0-35B-A3B-preview model card
huggingface.co
- 4NeoHorse-1-4B model card
huggingface.co
- 5DeepSeek-V4.1-Flash model card
huggingface.co
- 6YuE2-3B model card
huggingface.co
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.