🤗 Hugging Face Trending Models 2026-04-18

[github-actions]

Hugging Face Trending Models Digest 2026-04-18

Source: Hugging Face Hub | 30 models | Generated: 2026-04-18 00:13 UTC

---

Hugging Face Trending Models Digest — April 18, 2026

---

1. Today's Highlights

Google's Gemma 4 family continues to dominate the ecosystem with three variants in the top 30, collectively amassing over 8.5 million downloads. The most striking trend is the explosive community activity around uncensored and abliterated fine-tunes of Gemma 4, with four such models trending this week—suggesting strong demand for unfiltered open-weight alternatives. Qwen 3.6 is gaining serious momentum as a multimodal MoE architecture, with both official releases and community distillations like Jackrong's Claude-reasoning variant performing exceptionally well. Meanwhile, MiniMax M2.7 has emerged as a surprise challenger in the text-generation space, outperforming many larger models in engagement despite its compact size. The audio generation frontier is also heating up, with k2-fsa/OmniVoice nearing 850K downloads for zero-shot voice cloning.

---

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	1,381	100,019	Next-gen GLM with MoE+DSA architecture, strong reasoning capabilities in a chat-optimized format
MiniMaxAI/MiniMax-M2.7	MiniMaxAI	923	188,737	Compact yet powerful Chinese-developed LLM punching above its weight class in conversational quality
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled	Jackrong	2,723	579,463	Community distillation transferring Claude-4.6-level reasoning into open Qwen weights—highest likes this week
Jiunsong/supergemma4-26b-uncensored-gguf-v2	Jiunsong	381	53,781	Popular uncensored GGUF of Gemma 4 for local inference, optimized for llama.cpp
LilaRest/gemma-4-31B-it-NVFP4-turbo	LilaRest	246	74,404	NVIDIA FP4-quantized Gemma 4 variant pushing efficiency boundaries for RTX 50-series GPUs
OBLITERATUS/gemma-4-E4B-it-OBLITERATED	OBLITERATUS	275	6,866	Aggressively abliterated Gemma 4 E4B with both Safetensors and GGUF formats for maximum flexibility

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

Model	Author	Likes	Downloads	Why It's Trending
google/gemma-4-31B-it	google	2,122	3,513,465	Flagship vision-language Gemma 4 with massive adoption—over 3.5M downloads in its first weeks
Qwen/Qwen3.6-35B-A3B	Qwen	724	21,180	35B MoE with only 3B active params, delivering efficient image-text-to-text with strong benchmark scores
tencent/HY-Embodied-0.5	tencent	852	1,287	Hunyuan's first embodied AI vision model for robotics and spatial reasoning applications
baidu/ERNIE-Image	baidu	422	2,254	Apache-2.0 licensed 8B text-to-image model from Baidu, competitive with mid-tier SD variants
baidu/ERNIE-Image-Turbo	baidu	289	2,992	Faster inference variant of ERNIE-Image with minimal quality trade-off
tencent/HY-World-2.0	tencent	214	0	Ambitious world model for image-to-3D generation, fresh release with zero downloads but high anticipation
netflix/void-model	netflix	866	0	Netflix's video-to-video inpainting model for object removal—surprising corporate open-weight release
k2-fsa/OmniVoice	k2-fsa	629	847,404	Near-million-download zero-shot multilingual TTS with voice cloning, approaching mainstream adoption
openbmb/VoxCPM2	openbmb	1,095	18,089	Research-grade multilingual TTS with exceptional prosody control and emotional range
google/gemma-4-E4B-it	google	723	1,950,853	Experimental "any-to-any" Gemma 4 variant—unified encoder-decoder for arbitrary modality interconversion
google/gemma-4-26B-A4B-it	google	726	2,562,020	Mid-size vision-language Gemma 4 with 4B active params in MoE, optimal compute-quality balance
NucleusAI/Nucleus-Image	NucleusAI	158	802	Sparse MoE diffusion architecture demonstrating parameter-efficient image generation
LiconStudio/Ltx2.3-VBVR-lora-I2V	LiconStudio	135	5,539	Specialized LoRA for logical video reasoning in image-to-video generation, niche but innovative

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

Model	Author	Likes	Downloads	Why It's Trending
nvidia/Lyra-2.0	nvidia	161	35	NVIDIA's research artifact for arXiv:2604.13036—likely embedding or retrieval model, details sparse
OpenMOSS-Team/MOSS-TTS-Nano-100M	OpenMOSS-Team	132	29,102	Ultra-compact 100M Chinese TTS for edge deployment and resource-constrained environments

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

Model	Author	Likes	Downloads	Why It's Trending
dealignai/Gemma-4-31B-JANG_4M-CRACK	dealignai	1,257	152,762	MLX-native abliterated Gemma 4 for Apple Silicon, "4M-CRACK" suggests extensive fine-tuning dataset
unsloth/Qwen3.6-35B-A3B-GGUF	unsloth	363	153,019	Unsloth's optimized GGUF conversion enabling local multimodal MoE inference—153K downloads shows demand
Jiunsong/supergemma4-26b-uncensored-mlx-4bit-v2	Jiunsong	170	11,971	Apple Silicon-optimized 4-bit MLX of uncensored Gemma 4, filling a specific local inference niche
unsloth/ERNIE-Image-Turbo-GGUF	unsloth	150	13,664	Experimental GGUF quantization of diffusion model—unusual application of format typically used for LLMs
unsloth/MiniMax-M2.7-GGUF	unsloth	128	78,293	Rapid community quantization of trending MiniMax model, 78K downloads in short timeframe
HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive	HauhauCS	387	656,354	Most-downloaded uncensored variant this week—"Aggressive" abliteration with vision capabilities preserved
unsloth/gemma-4-26B-A4B-it-GGUF	unsloth	523	2,203,787	Over 2.2M downloads—likely the most successful quantization release ever, making MoE VLMs accessible
Qwen/Qwen3.6-35B-A3B-FP8	Qwen	108	22,905	Official FP8 release for H100/B200 deployment, showing Qwen's enterprise infrastructure focus

---

3. Ecosystem Signal

Gemma 4 has achieved ecosystem dominance unmatched since Llama 2's release, with Google's strategic release of three tiered variants (E4B "any-to-any," 26B-A4B MoE, 31B dense) creating a complete product ladder. The sheer volume of community fine-tunes—particularly uncensored and abliterated versions—demonstrates both the model's architectural accessibility and unmet demand for unfiltered open weights. This represents a significant shift: Google is successfully cultivating a developer community around its open models that rivals Meta's Llama ecosystem.

Quantization has become a primary distribution channel, not merely an optimization. Unsloth's GGUF conversions of Gemma 4 and Qwen 3.6 are achieving download volumes exceeding many base models, with their gemma-4-26B-A4B-it-GGUF nearing 2.2M downloads. This suggests local inference is transitioning from hobbyist niche to mainstream deployment pattern.

Chinese labs are increasingly competitive at the frontier. MiniMax, Qwen, Baidu, and Tencent all have trending releases this week, with Qwen 3.6's MoE architecture and MiniMax's efficiency gains challenging Western assumptions about open-weight leadership. The Jackrong distillation achieving the week's highest likes (2,723) also reveals sophisticated community knowledge transfer—distilling proprietary Claude reasoning into open Qwen weights.

Proprietary-to-open distillation and uncensored fine-tuning are the two most active community innovation vectors, sometimes converging as with the aggressive abliteration trend. Meanwhile, corporate open releases from unexpected sources (Netflix, Baidu's Apache licensing) suggest open weights are becoming standard competitive practice even for traditionally closed industries.

---

4. Worth Exploring

Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled — Highest engagement this week (2,723 likes, 579K downloads) for good reason: this represents a new paradigm in capability transfer, distilling frontier proprietary reasoning into an efficient open model. Worth studying for both the technical methodology and as a practical powerful reasoning model. The combination of Qwen's strong base with Claude-level chain-of-thought quality at 27B parameters challenges assumptions about necessary model scale.

google/gemma-4-E4B-it — The "any-to-any" experimental variant is the most architecturally ambitious release this week. With nearly 2M downloads despite its research status, it's worth exploring for understanding Google's multimodal unification strategy. If the "any-to-any" claims hold, this could presage a fundamental shift away from pipeline-based multimodal systems toward unified architectures—relevant for anyone building cross-modal applications.

k2-fsa/OmniVoice — Approaching mainstream adoption with 847K downloads, this zero-shot multilingual TTS with voice cloning achieves quality previously requiring proprietary APIs. The k2-fsa team's speech processing expertise shows in the natural prosody and cross-lingual transfer. Worth trying for any voice application, and particularly valuable for studying how open TTS is approaching human parity in 2026.

---

This digest is auto-generated by agents-radar.