HF breakouts Jul 13–20: Inkling clears the strict screen (2026)

The strict screen for Jul 13–20 has one clean, directly verified >10x breakout: thinkingmachines/Inkling. Its rolling 30-day download snapshot moved from 4 on Jul 16 to 12,456 on Jul 17, or more than 3,100× in adjacent daily snapshots. 1 2

That number needs a precise reading: Hugging Face exposes a rolling 30-day count, not a clean cohort-level weekly series. The spike is therefore best treated as a release-and-discovery signal, not as a claim that exactly 12,452 downloads arrived in one day. The rest of this issue is a watchlist rather than forced inclusions: Bonsai, MOSS-Transcribe-Diarize, and Wan-Dancer all have real traction, but the available snapshots do not establish a comparable >10x baseline.

Strict screen and watchlist

Model	Modality	Verified traction signal	License	Builder read
thinkingmachines/Inkling	Multimodal LLM	4 → 12,456 rolling 30-day downloads in adjacent Jul 16–17 snapshots; >3,100×	Apache-2.0, as listed in the daily screen 1	Evaluate for multimodal agents and document or audio workflows, but use hosted inference or a serious GPU budget rather than assuming laptop deployment.
prism-ml/Ternary-Bonsai-27B-gguf	Local text generation	301,893 rolling 30-day downloads on Jul 17; no comparable prior snapshot found	Apache-2.0 3	Strong watchlist candidate for private, laptop-friendly inference; not counted as a strict breakout this week.
OpenMOSS-Team/MOSS-Transcribe-Diarize	Audio ASR + diarization	86,385 rolling 30-day downloads on Jul 17; 87,533 on the current model card	Apache-2.0 4	A practical self-hosted meeting and podcast transcription option; the strict >10x test is unproven from the available snapshots.
Wan-AI/Wan-Dancer-14B	Music-to-video	2,328 rolling 30-day downloads on Jul 17; 2,408 on the model card	Apache 2.0 5	Interesting for character and dance-video products, but still a research deployment rather than a turnkey production dependency.

Inkling: the week’s only clean pass

Thinking Machines published Inkling on Jul 15 as its first open-weight model. It accepts text, images, and audio and produces text, using a sparse mixture-of-experts architecture with 975B total parameters and 41B active parameters. The model card also lists local and hosted paths through Transformers, SGLang, vLLM, llama.cpp, Unsloth, and Hugging Face Inference Providers. 6 7

The commercial case is unusually clear for a model this large. The current screen labels Inkling Apache-2.0, so commercial use is permitted under that license, subject to the model’s separate acceptable-use and safety requirements. 1 6

The catch is deployment cost. The official release notes put the BF16 checkpoint around 2 TB of VRAM and the NVFP4 variant around 600 GB, while also pointing to a 1-bit quantized path. 7 That makes Inkling a good candidate for an inference-provider evaluation, a well-funded self-hosted stack, or a product where multimodal breadth matters more than per-request simplicity. It is not a sensible default for an indie developer who expects to run the base checkpoint on a single workstation.

A practical evaluation should focus on the parts that differentiate Inkling from smaller multimodal models: audio understanding, long-context retrieval, tool use, and the quality of its responses when image and audio evidence arrive in the same task. The public model card reports its own benchmark results, but those results should be treated as a starting point for a task-specific eval rather than a production guarantee. 6

Watchlist: local models are getting the attention

Ternary Bonsai 27B: a useful local-inference bet

Bonsai is the most commercially practical item in the watchlist. The Hugging Face card describes a derivative of Qwen3.6-27B using ternary weights in GGUF Q2_0_g128 format, with an effective storage cost of about 1.71 bits per weight and a deployed size of roughly 7.2 GB. It runs through llama.cpp on CUDA, Metal, or CPU, and the card reports about 26 tok/s on an Apple M5 Pro. 3

That combination is exactly what many small teams want: a model that can stay inside a private environment and fit a consumer-oriented local stack. The Apache-2.0 license also avoids the commercial ambiguity that has blocked several recent breakouts. The important caveat is evidentiary, not technical: the Jul 17 screen showed 301,893 rolling 30-day downloads, and the Jul 19 screen showed 339,000, but no comparable pre-spike snapshot was available. 2 8 Put it in the eval queue, not in the >10x count.

MOSS-Transcribe-Diarize: one-pass speech workflows

MOSS-Transcribe-Diarize compresses a common product pipeline into one model: transcription, timestamps, and relative speaker labels in a single pass. Its model card says it supports 50+ languages, recordings up to 90 minutes, custom hotwords, and acoustic-event annotations. It is a 0.9B-parameter Apache-2.0 model, with serving paths through SGLang Omni and vLLM, although the card notes runtime caveats around CUDA 12 and the need for a pinned nightly vLLM build. 4

For builders, the opportunity is straightforward: private meeting notes, interview archives, podcast post-production, and call-center QA without stitching a separate diarization service onto ASR. The current 30-day count is healthy, but the available Jul 17 and Jul 19 snapshots move from 86,385 to roughly 87,500, nowhere near a verified tenfold jump. 2 8

Wan-Dancer-14B: a focused creative-video model

Wan-Dancer is a narrower bet: it generates long-duration dance videos from music, using a global keyframe-planning stage followed by local temporal refinement. The project released weights and inference code on Jul 13, and the model card links a project page, GitHub repository, Hugging Face Space, and ModelScope demo. It is released under Apache 2.0. 5

This is relevant to character animation, music promotion, and short-form video tooling, but the setup is heavier than the download count suggests: the repository asks users to install a custom stack, download the 14B checkpoint, and run separate global and local generation stages. The model card reports 2,408 downloads last month, up only slightly from the Jul 17 screen’s 2,328. 5 2 Treat it as a creative prototype candidate, not this week’s breakout.

What to build this week

Need broad multimodal reasoning and can pay for inference? Start with Inkling. Test image-plus-audio tasks, tool use, and long-context retrieval before committing to its infrastructure footprint.
Need private local inference? Start with Ternary Bonsai. Compare its quality and latency against the non-quantized base on your own prompts; the 7.2 GB footprint is the main practical advantage.
Need meeting or podcast automation? Test MOSS-Transcribe-Diarize on noisy recordings, overlapping speakers, and domain vocabulary. Do not treat [S01] as a real identity; the card defines those as relative labels. 4
Need music-synchronized creative video? Put Wan-Dancer in a sandbox and measure temporal continuity, render cost, and failure recovery before exposing it to customers.

Method and carry-over check

This screen covers the Jul 13–20 run window and uses rolling 30-day Hugging Face download snapshots from the daily trend reports. A model is counted as a strict breakout only when two comparable snapshots support a greater-than-tenfold increase. Inkling clears that test; the other three models have strong current traction but no defensible baseline in the retrieved window, so they remain watchlist entries rather than padded inclusions.

That distinction also prevents this issue from repeating the previous week’s names without new evidence. ThinkingCap and Krea identity-edit were already covered in the Jul 6–13 issue; the current screen does not provide a new >10x series that would justify promoting them again.

For an indie builder, the conclusion is less about chasing a long list than choosing the right constraint: Inkling for multimodal breadth, Bonsai for local efficiency, MOSS for speech workflows, and Wan-Dancer for a focused creative experiment.

Download counts are snapshots, not unique-user counts, and model-card claims should be validated on the target workload before production use.

HF breakouts Jul 13–20: Inkling clears the strict screen