Gemma 4 wins HF breakout week

From Jul 6 09:28 to Jul 13 09:00 PT, the Hugging Face breakout screen has one clean general-purpose winner: Google Gemma 4. Google DeepMind released five Gemma 4 open-weight sizes on Hugging Face on Jul 2, and the family reached roughly 2.77M downloads across its first 11 days with Apache 2.0 licensing across the set. 1

The rest of the week is a license and deployment filter. NuExtract3 has the biggest non-LLM download number, but its commercial status needs confirmation. 2 Cohere's Arabic speech model and NVIDIA's streaming automatic speech recognition (ASR) model are commercially usable audio picks, while Audex is research-only because NVIDIA released it under a non-commercial license. 345

Gemma 4 is the broadest commercial candidate because it combines a permissive license, multiple sizes, and multimodal inputs. The 12B Unified model is the most interesting architecture bet: Google DeepMind says it removes dedicated vision and audio encoders, then projects raw image patches and audio waveforms directly into the language-model embedding space through lightweight linear layers. 1

For a small team, that matters more than another leaderboard claim. A single family now covers small effective-parameter models, a 12B encoder-free multimodal model, a 26B A4B mixture-of-experts model, and a 31B dense model under the same Apache 2.0 umbrella. 1910 That gives builders a practical ladder: prototype on the smaller variants, then move up only if the task needs more reasoning or multimodal capacity.

ThinkingCap is narrower but actionable. BottleCapAI says the fine-tune was designed to preserve Qwen3.6-27B's answer quality and style while making it more token-efficient, and the model card reports a 45.8% average reduction in thinking tokens with a 0.8 percentage-point accuracy drop across out-of-domain benchmarks. 6 If your app pays for long reasoning traces, the relevant test is not only answer quality. The relevant test is whether shorter traces lower latency or inference cost without damaging your own task eval.

Cohere's Arabic ASR model is the cleanest new audio opportunity. The model card describes a 2B Conformer encoder-decoder model optimized for Arabic, Arabic dialects, English, and Arabic-English mixed speech. 3 Apache 2.0 makes the licensing easy, and the Open Universal Arabic ASR Leaderboard result gives teams a starting benchmark for Arabic transcription products. 3

Nemotron 3.5 ASR is not a fresh Jul 6 launch, but it belongs in the commercial audio shortlist. NVIDIA says the 600M streaming model can handle 240 concurrent streams at 80 ms latency on a single H100 GPU, compared with 14 streams for traditional buffered streaming inference. 4 That is a unit-economics claim worth testing if your product lives in live captions, meeting transcription, or voice-agent input.

NuExtract3 has the document-automation pull. The model card positions it for structured JSON extraction and image-to-Markdown conversion, and the current monthly download count is high enough to justify an evaluation pass. 2 The commercial decision should wait until the model page states license terms clearly. A document-ingestion startup should not build customer workflows on a model whose rights are still unresolved.

Kimi-K2.7-Code-NVFP4 is a deployment variant, not a new base model. NVIDIA's card says the quantization targets Blackwell GPUs, uses NVIDIA Model Optimizer v0.44.0, and runs through vLLM with tensor parallelism. 7 NVIDIA reports NVFP4 performance close to or slightly above INT4 baselines on SciCode, τ²-Bench Telecom, Terminal Bench 2.1, and SWE-bench Verified. 7 The builder question is simple: if you do not operate Blackwell inference, this is not where to spend the week.

Krea-2 identity-edit is useful for a smaller class of image products. The LoRA supports instruction-based edits that preserve a person or subject while changing scene, style, or local details. 8 The catch is the Krea 2 commercial threshold: the license allows commercial use only while the user and affiliated entities stay below $1M in annual revenue. 8 That can work for indie tools and creator workflows, but it is a bad foundation for a company planning to grow past the cap.

Audex is technically broad and commercially blocked. NVIDIA describes Audex-2B as a unified audio-text LLM, and the larger Audex-30B-A3B variant uses a 30B total-parameter, 3B-active mixture-of-experts (MoE) architecture with a 1M-token context window. 511 Both Audex releases use NVIDIA's Oneway Noncommercial License, so they should stay in research, demos, or competitive monitoring. 511

Three carry-over names should not be upgraded into commercial picks. Qwythos remains a legal-review item because its model card describes training on Claude Mythos and Claude Fable traces, even though the page lists Apache 2.0. 12 Sulphur-2-base remains out of commercial recommendations because the model page did not provide a clear license. 13 openPangu-2.0-Pro is also excluded until an official Pro model appears on Hugging Face; community previews are not a substitute for the promised official release.

The practical order for this week is: start with Gemma 4 if you want a general open-weight platform, test Cohere Arabic ASR or Nemotron ASR if speech is your product surface, and evaluate ThinkingCap if reasoning-token cost shows up in your margins. NuExtract3, Krea identity-edit, Kimi NVFP4, and Audex all have useful signals, but each one carries a specific blocker that should be resolved before it becomes production infrastructure.