HF breakouts Aug 17–24: no verified >10x jump; DeepSeek V4 Pro and Ornith 1.5 on watch (2026)

The evidence available through August 24, 2026 at 09:00 PT contains no Hugging Face model with a verified more-than-10x increase in comparable rolling 30-day downloads. The two new LLM candidates with the clearest movement are DeepSeek-V4-Pro-0813 and Ornith-1.5-35B-A3B: their observed increases are 1.76x and 1.38x. The metric is the Hub's rolling 30-day download count, so it signals activity on Hugging Face rather than an exact weekly cohort or a unique-user count. 1 2 3

What the screen found

The screen compares the same Hugging Face repository at two dated observations. A model enters the breakout list only when the later count is more than ten times the earlier count. The largest new signals in this week's evidence set remain below that bar.

Model	Modality	Comparable rolling 30-day downloads	Observed change	License and commercial status	Builder fit
DeepSeek-V4-Pro-0813	LLM	Aug 18: 30,985 → Aug 22: 54,566 1 3	1.76x (+76.1%)	MIT; generally permits commercial use under its terms. The model card lists no additional commercial restriction. 4	Agentic coding and tool use, with a very large serving footprint.
Ornith-1.5-35B-A3B	LLM	Aug 21: 9,165 → Aug 22: 12,611 2 3	1.38x (+37.6%)	The model card does not expose a license or commercial-use terms. Treat commercial status as unresolved. 5	Coding and terminal agents, with 3B active parameters but a multi-GPU serving requirement.

A higher download count would change the ranking, but it would not change the threshold. Both models belong on a builder's watchlist; neither is a verified breakout for August 17–24.

DeepSeek V4 Pro: commercial terms are clear, serving is not cheap

DeepSeek-V4-Pro-0813 is a 1.7-trillion-parameter model aimed at agentic workloads and production coding agents. The model card provides examples for vLLM and SGLang with expert parallelism, FP8 key-value caching, and speculative decoding. Those commands describe a multi-GPU deployment rather than a local developer setup. 4

The MIT license gives a startup a comparatively simple commercial path. The license still needs to be read alongside the model's code, dependencies, and deployment environment, but the Hub card does not add a separate revenue threshold or attribution condition. 4

The product question is whether agent quality pays for the infrastructure. A useful first test is a fixed set of coding tasks with tool calls, repository edits, and unit tests. Track end-to-end task success, incorrect tool calls, p95 latency, GPU hours, and cost per accepted change. The 1.76x download increase makes that test timely; it does not replace the test.

Ornith 1.5: a small active path with an unresolved license

Ornith-1.5-35B-A3B is a Mixture-of-Experts coding model with about 35B total parameters and about 3B active parameters per token. The card describes tool calling, agentic coding, and terminal coding agents. The card also lists a maximum 262,144-token context window and says that serving that context can use two 80GB GPUs. 5

The active-parameter count makes Ornith interesting for inference-cost experiments, while the full BF16 checkpoint still carries a roughly 70GB memory requirement. The model supports vLLM and SGLang with OpenAI-compatible serving and tool-call parsers. 5

The missing license is the gating issue. A startup can benchmark Ornith in a private research environment, but the current card does not support a commercial-use recommendation. Before a customer-facing prototype, obtain the repository's license terms and check whether the weights, code, and any derivative checkpoint carry the same permission.

Empty modality buckets

No new image-generation, audio, or multimodal candidate in the reviewed trend set has a paired in-window observation above the strict threshold. The evidence set is therefore LLM-only this week. The empty buckets are part of the result: a crowded trending list is not the same as a verified weekly breakout in every modality.

What to test this week

Run DeepSeek V4 Pro only where multi-GPU serving is available. Compare it with the current coding-agent baseline on the same repositories and tool permissions. Record successful changes, failed tool calls, latency, and cost per accepted change. 4
Keep Ornith 1.5 in a private evaluation lane. Test short coding tasks first, then test long-context tasks on the model card's two-80GB configuration. Hold commercial launch decisions until the license is visible and reviewed. 5
Wait for another dated snapshot before calling either model a breakout. The next observation should use the same repository and the same rolling-download definition, so the ratio remains comparable.

DeepSeek V4 Pro is the new candidate with the clearest commercial terms, but its 1.7T footprint makes serving the main product constraint. Ornith 1.5 offers a smaller active path for coding agents, while its unresolved license keeps it in private evaluation. Neither model clears the channel's strict more-than-10x screen by the August 24 cutoff.