HF breakouts Aug 10–17: LTX-2.5 clears 10x; three watchlist leads (2026)

Snapshot cutoff: August 17, 2026 at 09:00 Pacific Time. The comparison window runs from August 10 at 09:00 through the cutoff. This week's strict screen produces one verified >10x jump: Lightricks/LTX-2.5 moved from 39 to 378,439 displayed rolling 30-day downloads between the August 12 and August 15 snapshots. That is 9,703.6x. 1 2

The metric is a rolling 30-day repository-download count, not a count of people or an exact weekly cohort. Hugging Face says its download statistics count requests to repository files and do not deduplicate users. 3

The strict screen

Model	Modality	Earlier snapshot	Later snapshot	Derived growth	Result
Lightricks/LTX-2.5	Multimodal audiovisual generation	39, Aug 12	378,439, Aug 15	9,703.6x (378,439 ÷ 39)	Verified >10x breakout 1 2

The two observations use the same field, Downloads (30d), on the same Hugging Face trend surface. Both dates fall inside this week's window. No second comparable snapshot in the retrieved evidence supports a >10x result for an LLM-only, image-generation, or audio model.

LTX-2.5: a real spike with a real deployment bill

LTX-2.5 is an open-weight joint video-and-audio generator. Its model card lists text-to-video, image-to-video, video-to-video, audio-to-video, text-to-audio-video, and related input/output combinations. The model is built around a 22B diffusion transformer with separate video and audio components, and its headline use is synchronized video and audio generation from text, images, and video. 4

The practical product shape is short-form creative generation: ads, social clips, storyboard-to-video prototypes, image animation, and workflows that need sound attached to the clip. Lightricks also says the new release supports multishot generation, keeping character identity, environment, lighting, voice, and style across connected shots. That is more useful for a production prototype than a single impressive sample, but the result still needs a fixed prompt-and-asset evaluation set before it becomes a reliable feature. 4

The license is commercially usable with a revenue threshold. The card says companies under $10 million in annual revenue may use the model for commercial and production purposes at no cost under the LTX-2.x Community License. Companies above that threshold need a paid commercial-use agreement, and revenue is measured across the entity's subsidiaries and affiliates. Fine-tune transfer may also require a paid license. 4

Deployment is the catch. The card says the repository is not currently deployed by an Inference Provider. It offers ComfyUI, Diffusers, and an LTX pipeline path, plus 17 quantizations and five Spaces in the model tree. The full pack is split across a 22B DiT, a Gemma 4 12B text encoder, video and audio VAEs, and other components. A small team can prototype locally, but it should budget for storage, GPU time, and workflow integration rather than assume a one-click hosted endpoint. 4

The traction clears the channel's threshold; it does not prove output quality, unit economics, or durable adoption. The first builder test should compare LTX-2.5 with the current video baseline on prompt adherence, shot continuity, generation time, audio-video sync, and cost per accepted clip.

Watchlist by modality

These entries have a dated August 15 traction snapshot and a useful model card, but the evidence set lacks a paired prior snapshot that would justify a strict ratio. Their counts are signals to test, not breakout claims.

Model	Modality	Aug 15 rolling 30-day downloads	License and commercial status	Builder use
Qwen/Qwen3.8-27B	Multimodal: image/video understanding plus text	91,917	Apache-2.0; commercial use is generally permitted under the license	A 27B vision-language agent for documents, screenshots, long videos, coding, and long-horizon workflows; serves through Transformers, vLLM, SGLang, or TokenSpeed. 2 5
meta-models/Muse-Glimmer-30B	Multimodal: image and text in, text out	246,454	Apache 2.0; the card states commercial and research use	A local coding or document agent with tool use, failure recovery, and screenshot understanding; 4-bit weights target a 24–32 GB VRAM envelope. 2 6
MiniMaxAI/MiniMax-Music3	Audio: text-to-music	5,079	MiniMax-Music3 Community License; commercial products must display the model name, and products above $20 million in aggregate annual revenue need prior written authorization	A self-hosted music-generation component for full songs, drafts, game audio, and creator tools; the model card says it can generate up to five minutes of 32 kHz stereo audio. 2 7 8

Qwen3.8-27B: the practical multimodal bet

Qwen3.8-27B is a 27B dense vision-language model that accepts images and videos and supports a native 262,144-token context, with an extension path to 1 million tokens. Its card positions it for coding, professional work, research, and long-horizon agentic tasks. The model card reports compatibility with Transformers, vLLM, SGLang, and TokenSpeed, and lists Qwen Cloud as a future hosted option. 5

The deployment signal is unusually strong for a watchlist model: at fetch time the Hub page showed 98,000 likes, 125 community entries, 14 Spaces, and 505 quantizations. Those figures indicate an active integration surface, not proof that the model will win on a specific workload. Apache-2.0 removes the license barrier that blocks many frontier models, but a 27B checkpoint still needs a serious GPU or a managed serving layer. 5

A builder should test it on document extraction, screenshot-grounded operations, and agent tasks that exceed a small model's context. The comparison should include the cost of the full context window, because a model that can accept a million tokens can also make careless prompts expensive.

Muse Glimmer: the cleanest local-agent candidate

Meta published its official Muse Glimmer announcement on August 10, 2026. The model is a 30B multimodal agent built for local use, with day-one support in Transformers, llama.cpp, vLLM, and Hugging Face Inference Endpoints. The model card describes a dedicated perception encoder, tool use, multi-step reasoning, failure recovery, and input from text plus images. 6 9

The card's 4-bit release is designed to fit within 24 GB or 32 GB of VRAM, with the language model compressed to under 20 GB. It reports 29.6B total parameters, a 131,072-plus context length, and 15 Spaces, 25 finetunes, and 134 quantizations in the model tree. The Hub page showed 1,650 likes and 56 community entries at fetch time. 6

Apache 2.0 and the explicit commercial-use statement make the legal path clearer than the custom-licensed entries. The limitation is scope: the model card says audio input and output are unsupported, video is processed as individual frames, and the model can still fail on unfamiliar multi-step tasks. It fits a private document assistant, coding agent, or screenshot-driven operations tool better than a voice or video product. 6

MiniMax Music 3: audio traction, custom terms

MiniMax Music 3 combines an 8B global language model, a 0.6B local language model, and a 2.4B flow-matching synthesis component. It generates complete songs up to five minutes with lyrics and a structured music description, and outputs 32 kHz, 16-bit stereo WAV. The card recommends SGLang-Omni, Diffusers, and ComfyUI; it also says full precision fits under 24 GB of VRAM, while CPU offloading can reduce the requirement to about 22 GB and layer streaming can fit some 8 GB cards. 7

The license is the product constraint. Commercial interfaces must prominently display MiniMax-Music3. If the aggregate annual revenue from the relevant products and services exceeds $20 million, the operator must obtain prior written authorization from MiniMax. Hosted generation also requires reasonable technical and organizational safeguards against prohibited access, uses, and outputs. The license permits use, copying, modification, and distribution subject to those terms and its Acceptable Use Policy. 8

The fetched Hub page showed 888 likes, 19 community entries, 12 Spaces, and 18 quantizations, but no Inference Provider deployment. The card's retrieved view does not expose a reliable first-published date, so the August 15 count is a new-signal observation rather than evidence of a multi-day growth ratio. A creator-tool team should first test genre consistency, vocal identity, lyric adherence, and output moderation before building a customer-facing generation flow.

Empty buckets are part of the result

The strict screen has no image-generation candidate. LTX-2.5 generates video and audio from images and text; it does not fill a text-to-image slot. The strict screen also has no audio candidate. MiniMax Music 3 is worth watching, but its August 15 observation has no paired earlier count in this evidence set. The LLM-only bucket is empty as well; Qwen3.8-27B and Muse Glimmer are multimodal models, so they stay in that section.

Hugging Face's summer report gives the broader reason to keep this discipline: downloads measure activity inside the Hub, while likes measure attention, and neither is a direct measure of model quality, commercial adoption, or market share. The report also finds that models below 1B account for 83% of all-time downloads, while models above 100B account for 1%, which is a reminder that frontier size alone is a poor deployment filter. 10

What a builder can do this week

Prototype the strict breakout only where audiovisual generation is the product. Start with LTX-2.5 on a fixed set of image-to-video prompts. Record accepted-clip rate, generation time, audio-video sync, and cost per accepted output. Check the revenue threshold before any paid launch. 4
Choose Muse Glimmer or Qwen3.8-27B for a local multimodal agent test. Use the same screenshots, documents, and tool-call harness for both. Measure successful end-to-end task completion rather than relying on benchmark rank. 5 6
Treat MiniMax Music 3 as a rights-and-quality experiment. Keep the model label in the interface, log the license-triggered revenue boundary, and test moderation and rights workflows before collecting user-generated prompts. 8

The useful conclusion is narrow: LTX-2.5 is the only verified >10x breakout in this week's evidence, and its spike comes with a self-hosting and licensing decision attached. Muse Glimmer and Qwen3.8-27B are the cleanest multimodal prototypes; MiniMax Music 3 is the audio lead, but its community license needs to be part of the product design from day one.

참고 출처

1
GenAI Secret Sauce Daily Digest — 2026-08-12genaisecretsauce.com
2
GenAI Secret Sauce Daily Digest — 2026-08-15genaisecretsauce.com
3
Hugging Face model download statisticshuggingface.co
4
Lightricks/LTX-2.5 model cardhuggingface.co
5
Qwen3.8-27B model cardhuggingface.co
6
Muse Glimmer model cardhuggingface.co
7
MiniMax Music 3 model cardhuggingface.co
8
MiniMax-Music3 Community Licensehuggingface.co
9
Meta is back with Muse Glimmerhuggingface.co
10
State of Open Models: Summer 2026 Observationshuggingface.co