HF breakouts Jul 27–Aug 3: DeepSeek V4 Flash-0731 clears 10x; six watchlist leads (2026)

Snapshot cutoff: Aug 3, 2026 at 09:00 (UTC-08:00). The comparison window starts Jul 27 at 09:00.

One model clears the strict screen

deepseek-ai/DeepSeek-V4-Flash-0731 is the only model in this week's evidence set that clears a strict >10x download-growth screen. A Jul 31 snapshot reported about 940 rolling 30-day downloads for the new build; the Hugging Face model card showed 236,076 when checked for this edition. That is an observed jump of about 251x. 1 2

The ratio is a traction signal, not a count of weekly users. Hugging Face counts server-side requests to selected repository query files, including GET and HEAD; it does not deduplicate users. The current card also does not expose an exact timestamp for its displayed count, so 251x means Jul 31 snapshot to the Aug 3 observation, not an exact seven-day cohort. 3

Model	Modality	Earlier rolling 30-day snapshot	Later snapshot	Observed change	Screen
deepseek-ai/DeepSeek-V4-Flash-0731	LLM	~940, Jul 31	236,076, Aug 3 observation	~251x	Verified >10x 1 2
moonshotai/Kimi-K3	Multimodal	387,822, Jul 30	837,000, Aug 2	2.16x	Watchlist 4 5
thinkingmachines/Inkling-Small	Multimodal	3,998, Aug 1	8,504, Aug 3 observation	2.13x	Watchlist 6 7
poolside/Laguna-S-2.1	LLM	67,300, Jul 29	77,021, Jul 31	1.14x	Watchlist 1 8
upstage/Solar-Open2-250B	LLM	4,800, Jul 29	14,900, Aug 2	3.10x	Watchlist 5 8
Kwaipilot/KAT-Coder-V2.5-Dev	LLM	6,280, Jul 29	10,771, Aug 1	1.72x	Watchlist 6 8
microsoft/VibeVoice-ASR	Audio	No comparable prior snapshot	694,766, Aug 3 observation	Not computable	Watchlist 9 10

The strict breakout: DeepSeek V4 Flash-0731

This is an official DeepSeek-V4-Flash release with a 304B-parameter architecture and a speculative-decoding module attached. The model card reports agentic coding and tool-use results including 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. The card also describes a 1M-token context and selectable reasoning effort. Those numbers are the publisher's reported evaluations, not an independent test of your application. 2

The license is MIT, with no extra commercial restriction stated on the card. Deployment is still far from lightweight: the card's vLLM example enables expert parallelism, data parallelism of four, FP8 KV cache, and DSpark speculative decoding; its SGLang example uses tensor parallelism of four. That makes the model interesting for a self-hosted coding agent or repository automation service, but not an obvious first experiment for a solo developer without multi-GPU access. 2

A sensible business test is narrow: run it against a fixed set of private repositories and measure patch acceptance, tool-call failure, latency, and cost per completed task. Do not treat the 251x Hub jump as evidence that it will beat a smaller coding model on your workload.

LLM watchlist

Laguna-S-2.1: the most deployment-friendly coding lead

Poolside's 118B MoE model has roughly 8B active parameters per token, a 1M-token context, and support for vLLM, SGLang, Transformers, TRT-LLM, llama.cpp, and Ollama-related workflows. The card lists FP8, NVFP4, INT4, and GGUF variants; the BF16 checkpoint is still roughly 236GB of weights. Its reported scores include 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-bench Multilingual. 11

The OpenMDW-1.1 terms explicitly allow commercial and non-commercial use, subject to Poolside's Acceptable Use Policy. The business case is a self-hosted coding or codebase-Q&A product where quantization matters more than frontier breadth. The download curve does not support a breakout claim this week; the signal is steady attention around a model with unusually clear deployment paths. 11

Solar Open 2: useful capabilities, unresolved license labeling

Solar-Open2-250B is a 250B/15B-active hybrid-attention MoE aimed at office productivity, document-heavy work, coding, tool calling, and multi-step reasoning. The card reports a 1M-token context and recommends vLLM. Its minimum hardware is four H200s, with eight recommended; official INT4 and NVFP4 quantized variants are available. 12

There is a material license conflict. The Aug 2 digest labels the repository Apache 2.0, check repo, while the model card identifies the Upstage Solar License and requires derivative names to use the Solar prefix, public materials to display Built with Solar, and the license to travel with the derivative. Until Upstage confirms which terms govern the exact files you will ship, treat Solar as a research or enterprise-evaluation candidate, not a clean commercial dependency. 5 12

KAT-Coder: strong coding shape, thin legal signal

KAT-Coder-V2.5-Dev is a text-only 35B MoE with 3B active parameters. The card reports 69.40 on SWE-bench Verified, 63.00 on SWE-bench Multilingual, and support for Transformers, vLLM, SGLang, and KTransformers. Its example serving configurations use eight-way tensor parallelism and a 262,144-token context. 13

The Aug 1 digest labels it Apache-2.0, but the model-card text retrieved for this edition does not state a license or commercial-use grant. That disagreement is enough to block a commercial recommendation until the repository metadata and the license file are checked together. For builders, the likely use is pull-request triage or a repository agent; the hardware requirement is the first practical filter. 6 13

Image generation: no verified breakout

No image-generation model met the two-snapshot >10x bar in this week's evidence. The image-related entries that did surface were OCR or multimodal models, which solve different problems. Keeping that bucket empty is more useful than presenting a newly visible image repository as evidence of traction.

Audio watchlist

VibeVoice-ASR: a real product lead, not a verified breakout yet

Microsoft's 9B VibeVoice-ASR handles up to 60 minutes of audio in one pass, jointly producing speaker labels, timestamps, and transcription. It supports more than 50 languages, code-switching, and custom hotwords. The card links to a vLLM integration and reports 23 Spaces using the model. The August trends collection added it on Aug 3 with 694,156 displayed downloads, while the model card showed 694,766; the 610-count difference between those two Hub surfaces does not change its watchlist status. 9 10

The license is MIT. Its practical business fit is long-form meeting, interview, podcast, or support-call transcription where speaker continuity matters. There is no comparable prior snapshot available, so its current popularity stays on the watchlist rather than entering the strict breakout count. 10

Multimodal watchlist

Kimi K3: frontier breadth with a legal and hardware bill

Kimi K3 is a 2.8T-parameter MoE with 104B active parameters, a 1M-token context, native vision, and support for long-horizon coding, research, tool use, and visual work. The card reports 21 Spaces, 32 finetunes, and 29 quantizations. It can be accessed through Kimi's API or served with vLLM, SGLang, or TokenSpeed. 14

Its license is named Kimi K3 License, but the card does not state an explicit commercial-use grant. The full model is also a serious hosting project. The week-over-week signal is attention, not breakout: 387,822 rolling 30-day downloads on Jul 30 rose to 837,000 on Aug 2, about 2.16x. A visual research assistant or long-context agent is a plausible experiment; a self-hosted product should begin with a quantized variant and a real cost test. 4 5 14

Inkling-Small: compact inputs, unresolved rights

Inkling-Small accepts text, images, and audio and emits text through a sparse MoE decoder. The card describes a 276B-total/12B-active architecture, shared processing across modalities, and support for agentic systems, coding assistants, chatbots, and retrieval-augmented generation. It reports 29 quantizations and three Spaces. 7

The Aug 1 digest labels it Apache-2.0, but the fetched card exposes an Acceptable Use Policy without naming a license or giving a commercial-use grant. Its displayed downloads rose from 3,998 on Aug 1 to 8,504 on the current card, about 2.13x. That is useful early traction for a smaller multimodal interface, but the rights question should be resolved before it enters a customer-facing stack. 6 7

What the numbers support

Test the verified breakout as a coding model, not as a generic foundation model. DeepSeek V4 Flash-0731 has the clearest traction jump, MIT terms, and a documented agent/tool-use path. Its multi-GPU serving cost is the counterweight.
Treat audio as the strongest non-LLM product lead. VibeVoice-ASR has a concrete workflow—long-form transcription with diarization and timestamps—and MIT terms, but it needs a prior snapshot before it can be called a breakout.
Do not convert label visibility into legal clearance. Solar Open 2, KAT-Coder, and Inkling-Small all show a gap between a digest's metadata label and the license or policy text visible on the model card. That gap is itself a deployment blocker.
Keep image generation out of this week's shortlist. No verified candidate cleared the screen, and a blank bucket is more informative than a weakly related recommendation.

For builders, this week's distinction is sharp: one model has a measured surge; the rest are leads with different combinations of capability, rights, and serving cost. The next useful step is a task-level benchmark on the model that fits your hardware and license boundary—not another read of the trending list.

References

1
GenAI Secret Sauce Daily Digest — Jul 31, 2026genaisecretsauce.com
2
DeepSeek-V4-Flash-0731 on Hugging Facehuggingface.co
3
Hugging Face model download statisticshuggingface.co
4
Jul 30 digestgenaisecretsauce.com
5
Aug 2 digestgenaisecretsauce.com
6
Aug 1 digestgenaisecretsauce.com
7
Inkling-Small model cardhuggingface.co
8
Jul 29 digestgenaisecretsauce.com
9
AI Trends 2026-08 collectionhuggingface.co
10
VibeVoice-ASR model cardhuggingface.co
11
Laguna-S-2.1 model cardhuggingface.co
12
Solar-Open2-250B model cardhuggingface.co
13
KAT-Coder-V2.5-Dev model cardhuggingface.co
14
Kimi K3 model cardhuggingface.co