HF breakouts Jul 20–27: no strict >10x model, but six builder leads (2026)

The strict screen found zero verified breakouts

For Jul 20–27, no model clears a strict >10x rule. The rule requires two comparable rolling-download snapshots, not a first appearance on a trend list. The overlapping models we could compare moved between roughly 1.2× and 2.7×, well below the threshold. 1

Model	Jul 19 rolling 30-day downloads	Latest observed Jul 27	Approx. change
zai-org/GLM-5.2	536,000	1,003,547	1.9× 2
baidu/Unlimited-OCR	2.12M	2,645,773	1.2× 3
prism-ml/Ternary-Bonsai-27B-gguf	339,000	648,938	1.9× 4
thinkingmachines/Inkling	13,500	36,196	2.7× 5
OpenMOSS-Team/MOSS-Transcribe-Diarize	87,500	125,294	1.4× 6
Wan-AI/Wan-Dancer-14B	2,400	4,121	1.7× 7

These are rolling 30-day download counts, not exact weekly cohorts or unique users. Hugging Face says its counter is based on requests to repository query files, so treat the numbers as a discovery and release signal rather than product adoption. 8

High-traction watchlist

The late-week candidates below are useful signals, but they are not being presented as verified >10x breakouts. The Jul 26 snapshot reported 30-day counts of 35,562 for OvisOCR2, 32,700 for Cosmos3-Edge, 5,978 for Antares-1B, and 3,764 for KAT-Coder-V2.5-Dev. A separate Jul 25 snapshot put Mage-Flow at about 1,200 and Solar-Open2-250B at about 2,800. 9 10

Multimodal and document systems

ATH-MaaS/OvisOCR2

OvisOCR2 is a 0.8B page-level document parser built by post-training Qwen3.5-0.8B. It turns a document-page image into Markdown in reading order, including text, formulas, tables, and visual regions. The model card reports an OmniDocBench v1.6 score of 96.58 and gives a vLLM deployment path. 11

License: Apache-2.0. The card does not add separate commercial-use language, so the license is the clearest permission signal rather than a substitute for checking data, trademark, and output obligations. 11
Business fit: Try it first for invoice, claims, forms, and contract-ingestion pipelines where the output needs to remain editable. Its small size and vLLM path make it the most straightforward candidate here for a document-processing prototype. Keep human review in the loop for legally or financially consequential fields; the card warns that outputs can be incomplete or incorrect.

nvidia/Cosmos3-Edge

Cosmos3-Edge is a 4B omnimodal world model for physical-AI work. The card describes inputs and outputs spanning text, images, video, audio, and action commands, with applications in robotics, autonomous vehicles, and smart spaces. The Jul 26 snapshot recorded 32,700 rolling 30-day downloads. 12 9

License: OpenMDW1.1; the card explicitly says it is ready for commercial and non-commercial use. 12
Business fit: It is a plausible starting point for simulation, robot-policy experiments, and edge-vision prototypes. Deployment is the constraint: the card lists Linux, NVIDIA Ampere/Blackwell/Hopper hardware, BF16 as the tested precision, and vLLM-Omni or vLLM. That is a serious infrastructure commitment for a small team.

LLMs for code, security, and agents

fdtn-ai/antares-1b

Antares-1B is a security-specialized 1B model built on IBM Granite 4.0 1B. It operates as a terminal agent that searches a repository to localize files likely to contain a vulnerability. The card reports a File F1 of 0.209 on 500 VLoc Bench tasks and a full sweep of about 13 minutes on one H100 with 16 parallel workers. 13

License and access: Apache-2.0 metadata is present, but the repository is gated and requires a manual access request. That makes it legally more promising than a non-commercial model, but operationally less frictionless than an openly downloadable checkpoint. 13
Business fit: Use it as a component in an internal CI/CD vulnerability-triage workflow, not as a general chat model. The recommended setup is a disposable, network-disabled sandbox with resource limits, and the card recommends human oversight.

Kwaipilot/KAT-Coder-V2.5-Dev

KAT-Coder-V2.5-Dev is a text-only MoE coding model with 35B total parameters and 3B active parameters, based on Qwen3.6-35B-A3B. The card lists agentic software-engineering evaluations and supports Transformers, vLLM, SGLang, and KTransformers. It also recommends a 262,144-token context window. 14

License and access: Hugging Face metadata labels the repository Apache-2.0, while the model-card text does not spell out commercial-use guidance. Treat that as a verify-before-shipping item, not a clean legal green light. 14
Business fit: The practical target is a self-hosted coding agent, pull-request triage, or repository Q&A. The text-only checkpoint avoids a common multimodal-loading trap, but the long-context serving configuration still needs a real memory and latency test.

upstage/Solar-Open2-250B

Solar Open 2 is a 250B-A15B agentic MoE model for office productivity, document-heavy work, tool calling, and coding. Its hybrid attention stack keeps 15B parameters active per token and supports a 1M-token context. The Jul 25 snapshot showed about 2,800 rolling 30-day downloads. 15 10

License: Upstage Solar License. Derivative models must use a name prefixed with "Solar," display "Built with Solar" in public-facing materials, and include the license. 15
Business fit: This is an on-premises enterprise experiment, not an indie-friendly first deployment. The card lists H200 ×4 as the minimum and H200 ×8 as the recommendation, plus a vLLM/CUDA 12.9 setup.

Image generation and editing

microsoft/Mage-Flow

Mage-Flow is a 4B-scale text-to-image and instruction-based image-editing stack. Its card reports 0.59 seconds per 1024² image for Mage-Flow-Turbo and 1.02 seconds per edit on a single A100, with peak memory around 18–20 GB. It supports native-resolution generation from 512 to 2048 and includes CLI and Gradio entry points. 16

License: MIT. The card does not add separate commercial-use terms beyond the license. 16
Business fit: It is the most approachable image candidate for branded creative tooling, catalog-image editing, or a private design assistant. Budget for CUDA and flash-attn setup work; the card specifically warns about matching the local CUDA toolkit to the PyTorch build.

What to build first

Document automation: benchmark OvisOCR2 on your own invoices and forms, measuring field-level accuracy rather than relying on the headline score.
Security triage: request Antares access and wire it into a disposable CI job with a human review step.
Creative tooling: test Mage-Flow-Turbo on your actual aspect ratios and typography before committing to a production GPU budget.
Physical AI: only take Cosmos3-Edge past a prototype if you already operate the required NVIDIA/Linux stack.
Defer the large-model bets: KAT-Coder and Solar Open 2 have interesting builder profiles, but this week’s evidence shows attention, not a verified tenfold weekly surge. For both, resolve licensing and serving cost before treating downloads as product demand.

For builders, the practical distinction is simple: a newly visible repository is a lead to test, not evidence of sustained adoption, until a second comparable rolling-download snapshot exists.

HF breakouts Jul 20–27: no strict >10x model, but six builder leads