🤗 Hugging Face Trending Models 2026-04-12

[github-actions]

Hugging Face Trending Models Digest 2026-04-12

Source: Hugging Face Hub | 30 models | Generated: 2026-04-12 00:12 UTC

---

Hugging Face Trending Models Digest — April 12, 2026

---

1. Today's Highlights

This week's trending models reveal Google's Gemma 4 family dominating the ecosystem with multiple variants spanning 2B to 31B parameters, including the experimental "any-to-any" E4B and E2B models. Community fine-tuning activity is exceptionally active, with HauhauCS releasing multiple uncensored variants and Jackrong distilling reasoning capabilities from Claude 4.6 Opus into Qwen 3.5 architectures. Notably, nvidia/personaplex-7b-v1 signals growing interest in speech-to-speech AI agents, while netflix/void-model represents a rare video-to-video entry from a major studio. The continued proliferation of GGUF quantizations by Unsloth underscores persistent demand for local, efficient inference.

---

2. Trending Models

🧠 Language Models (LLMs, chat models, instruction-tuned)

Model	Author	Likes	Downloads	Why It's Trending
zai-org/GLM-5.1	zai-org	985	23,998	Latest GLM MoE architecture with DSA optimization, representing Zhipu's continued push for efficient large-scale language models.
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled	Jackrong	2,582	566,643	High-engagement reasoning distillation combining Qwen 3.5 architecture with Claude 4.6 Opus reasoning patterns.
prism-ml/Bonsai-8B-gguf	prism-ml	560	71,661	Extreme 1-bit quantization pushing the boundaries of minimal viable LLMs for edge deployment.
LGAI-EXAONE/EXAONE-4.5-33B	LGAI-EXAONE	105	3,396	LG's latest Korean-optimized multimodal model, representing regional foundation model competition.
kai-os/Carnice-9b	kai-os	149	4,147	Hermes-agent tuned Qwen 3.5 variant optimized for autonomous agent workflows.

🎨 Multimodal & Generation (image, video, audio, text-to-X)

Model	Author	Likes	Downloads	Why It's Trending
google/gemma-4-31B-it	google	1,730	2,026,015	Flagship Gemma 4 vision-language model with massive adoption, setting the open-weight multimodal baseline.
google/gemma-4-E4B-it	google	587	1,136,791	Experimental "any-to-any" architecture enabling flexible input-output modality combinations.
google/gemma-4-E2B-it	google	405	774,725	Compact any-to-any variant democratizing multimodal AI for resource-constrained environments.
baidu/Qianfan-OCR	baidu	1,136	44,379	InternVL-based document understanding model targeting enterprise OCR and visual information extraction.
nvidia/personaplex-7b-v1	nvidia	2,422	473,689	Moshi-based speech-to-speech agent with persona capabilities, signaling NVIDIA's voice AI ambitions.
k2-fsa/OmniVoice	k2-fsa	500	340,361	Zero-shot multilingual voice cloning with strong open-source community backing.
openbmb/VoxCPM2	openbmb	697	5,722	Next-generation TTS from OpenBMB with improved prosody and multilingual coverage.
netflix/void-model	netflix	758	0	CogVideoX-based video inpainting for object removal, representing studio investment in generative video tooling.
tencent/HY-Embodied-0.5	tencent	126	535	Hunyuan vision-language model for embodied AI and robotics applications.
tencent/HY-OmniWeaving	tencent	247	0	Research preview of unified multimodal generation with dual arXiv preprints indicating active development.
LiquidAI/LFM2.5-VL-450M	LiquidAI	100	4,925	Ultra-efficient 450M parameter vision-language model from liquid neural network specialists.

🔧 Specialized Models (code, math, medical, embeddings)

Model	Author	Likes	Downloads	Why It's Trending
nvidia/Gemma-4-31B-IT-NVFP4	nvidia	352	565,972	NVIDIA-optimized FP4 quantization using ModelOpt, targeting H100/B200 deployment efficiency.

📦 Fine-tunes & Quantizations (community fine-tunes, GGUF, AWQ)

Model	Author	Likes	Downloads	Why It's Trending
dealignai/Gemma-4-31B-JANG_4M-CRACK	dealignai	930	89,753	MLX-compatible abliterated Gemma 4 for Apple Silicon, demonstrating strong community demand for uncensored local inference.
HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive	HauhauCS	265	373,322	Aggressive uncensored fine-tune of Gemma 4's any-to-any model, controversial but heavily downloaded.
HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive	HauhauCS	1,098	868,534	Most-liked uncensored Qwen 3.5 variant, indicating sustained demand for unfiltered local models.
HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive	HauhauCS	1,256	962,827	MoE-enabled uncensored vision model, combining efficiency and multimodal capabilities with controversial fine-tuning.
unsloth/gemma-4-26B-A4B-it-GGUF	unsloth	420	1,523,972	Highest-downloaded GGUF this period, Unsloth's efficient quantization enabling widespread local deployment.
unsloth/gemma-4-31B-it-GGUF	unsloth	281	997,865	Near-million download GGUF of Google's flagship, proving Unsloth's dominance in accessible quantization.
unsloth/gemma-4-E4B-it-GGUF	unsloth	236	932,802	Any-to-any model in GGUF format, extending multimodal capabilities to consumer hardware.
unsloth/GLM-5.1-GGUF	unsloth	115	21,864	Rapid GGUF conversion of newly released GLM-5.1, showing Unsloth's responsive ecosystem support.
Jackrong/Qwopus3.5-27B-v3-GGUF	Jackrong	264	111,740	Reasoning-optimized Qwen 3.5 in GGUF, community alternative to proprietary reasoning models.
LilaRest/gemma-4-31B-it-NVFP4-turbo	LilaRest	141	13,912	Community NVFP4 optimization pushing NVIDIA's format beyond official implementations.

---

3. Ecosystem Signal

Gemma 4 hegemony with experimental edges. Google's Gemma 4 family commands unprecedented ecosystem attention—8 of 30 trending models, spanning 2B to 31B parameters with three architectural tiers: standard vision-language (31B/26B), experimental any-to-any (E4B/E2B), and dense variants. The "any-to-any" pipeline tag represents Google's bid to define the next multimodal paradigm, though adoption lags behind proven image-text-to-text architectures.

Qwen 3.5 as the community's canvas. Alibaba's Qwen 3.5 architecture powers 5 trending models, predominantly through community fine-tunes. Jackrong's Claude distillation and HauhauCS's aggressive uncensored series demonstrate Qwen's emergence as the preferred base for experimental modifications—more tractable than Gemma for community modification, more capable than Llama derivatives for Asian-language contexts.

Quantization as infrastructure. Unsloth's systematic GGUF conversion of every major release—4 models this period—has created a de facto standard for accessible local inference. The download volumes (3.5M+ combined for Unsloth GGUFs) reveal a bifurcated ecosystem: researchers and enterprises use native PyTorch, while practitioners and hobbyists overwhelmingly prefer quantized formats.

Speech AI's inflection point. Three TTS/speech models trend simultaneously, with NVIDIA's personaplex-7b representing the most sophisticated entry: not merely text-to-speech, but speech-to-speech agents with persona persistence. This suggests 2026 as the year voice interfaces mature beyond transcription and synthesis toward genuine conversational agents.

The uncensored paradox. HauhauCS's three models combine 2,619 likes with 2.2M downloads—engagement ratios suggesting users download quietly what they hesitate to endorse publicly. This persistent demand creates tension with platform policies and model licenses, particularly as Gemma's terms explicitly prohibit certain modifications.

---

4. Worth Exploring

nvidia/personaplex-7b-v1 — This Moshi-based speech-to-speech model represents the most technically ambitious trending entry. Unlike conventional TTS, it maintains conversational state across voice exchanges with persona consistency. Worth studying for: (1) architectural patterns for real-time speech LLMs, (2) NVIDIA's emerging voice AI stack, (3) implications for human-computer interaction design. The 2,422 likes with 473K downloads suggest strong researcher interest despite limited public documentation.

Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled — The highest-liked model (2,582) demonstrates sophisticated knowledge transfer from proprietary to open-weight architectures. Worth trying for: evaluating whether distilled reasoning matches native capabilities, understanding distillation methodology for advanced reasoning patterns, and as a cost-effective alternative to API-dependent reasoning models. The 566K downloads indicate production deployment beyond experimentation.

google/gemma-4-E4B-it — The "any-to-any" experimental variant merits attention despite lower engagement than the standard 31B. Worth exploring for: architectural preview of Google's multimodal roadmap, testing novel input-output combinations (audio→image, video→text), and assessing whether flexible modality routing outperforms specialized pipelines. The 1.1M downloads suggest substantial curiosity, though practical applications remain undefined.

---

This digest is auto-generated by agents-radar.