HF breakouts Aug 24-31: zero verified >10x jumps; three models on watch

Evidence cutoff: August 31, 2026 at 09:00 PT. The paired-snapshot screen yielded zero verified >10x breakouts for August 24-31. A strict breakout requires two dated, comparable rolling-30-day download observations for the same Hugging Face repository. This week's evidence set supplied no such pair above 10x.

The three entries below are new watchlist signals, supported by in-window release or update evidence and current Hub activity. The Hub's Downloads last month field is a current discovery signal. It is not a weekly cohort, a unique-user count, or a growth rate. The absence of a paired snapshot leaves every watchlist entry without a defensible weekly multiplier.

Apodex-1.1-mini is a reasoning-first model for long-horizon research tasks. Its card describes work across papers, datasets, spreadsheets, images, code, and tools, with intermediate checking before delivery. 1 That combination gives a builder a defined product test: ask the model to turn a small folder of source files into a cited report, a cleaned data table, and a reproducible calculation.

The serving example uses SGLang or vLLM with tensor parallelism across eight devices and a maximum context of 262,144 tokens. 1 Apache 2.0 generally supports commercial deployment, but the license decision should cover the repository, model weights, code, and dependencies together. The first gate is operational: measure citation accuracy, structured-output validity, retry rate, latency, and hardware cost on the exact workflow a product would sell.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series, according to its Hub card. The card lists 320B total parameters and 18B active parameters, with support for image and video inputs through a vision encoder. 3 The release discussion dates the public release signal to August 26 and describes the FP8 and BF16 weight variants. 2

The license needs a manual check before a commercial recommendation. The Hub card's visible text leaves the license field unresolved, while the community release thread labels the model MIT. 23 The card lists SGLang, vLLM, Transformers, KTransformers, TokenSpeed, and Unsloth support. 3 A prototype should therefore start with document and image extraction on a controlled set, then pass only if the team can serve the full checkpoint within its latency and memory budget. The 18B active-parameter figure describes per-token activation; it does not turn the 320B checkpoint into an 18B download.

DeepSeek-V4-Flash-Vision-Exp adds visual modules to the DeepSeek-V4-Flash architecture for visual understanding. The model card calls it an experimental multimodal model and lists a 305B-parameter checkpoint. 4 The same card reports MIT licensing and says that no Inference Provider currently deploys the model. 4

A sensible prototype would use a fixed set of charts, scanned documents, and visual instructions. Record extraction errors, unsupported layouts, response time, and the cost of a second pass. The experimental status and absent hosted deployment make a fallback necessary before customer-facing use. MIT gives the repository a clear starting point for commercial review; it does not remove the cost of operating a 305B model or the need to review every dependency.

Image generation: zero new candidates met the paired-snapshot standard for August 24-31.

Audio: zero new candidates met the paired-snapshot standard for August 24-31.

Current Hub counts and community attention alone cannot fill either bucket because neither signal supplies the two dated observations needed for a strict >10x claim.

The evidence supports three prototypes worth testing. It supports no verified weekly breakout yet. A later dated Hub snapshot can turn any one of these watchlist entries into a strict comparison; this issue leaves the multiplier open until that observation exists.