← Back to Hugging Face models & datasets

Hugging Face models & datasets

The top 18 hugging face models & datasets measured by RiseFinder over the 30d window ending October 2026.

Period: Monthly
Window: 30d
Rows: 18
Data through: October 2026

← All current risers All hugging face models & datasets editions Older: September 2026

01/

Monthly box

18 entries

01

pipecat-ai/phonellm-alpha-1

PhoneLLM Alpha 1 is an open-weights language model from the Pipecat team designed for voice agent use cases, such as handling inbound customer service calls and outbound calling tasks when paired with transcription and text-to-speech models. It is a full-parameter fine-tune of NVIDIA's Nemotron 3 Nano 30B-A3B, a mixture-of-experts model with 3.5B active parameters, and is trained to invoke tools accurately in multi-turn conversations without requiring thinking tokens. It is released under the BSD license with no commercial restrictions and can be run on your own infrastructure.

Hugging Face model

Open source ↗

+499.9%

4,721 → 28,322 downloads

02

unsloth/GLM-5.3-Flash-GGUF

Unsloth/GLM-5.3-Flash-GGUF is a GGUF-format release of GLM-5.3-Flash, a large language model from the GLM-5 series. The model is described as the first natively multimodal model in that series, with 320B total parameters and 18B active parameters, using a hybrid architecture that combines sparse and linear attention to reduce long-context serving costs. The release is intended to be run with llama.cpp or the Unsloth Desktop app.

Hugging Face model

Open source ↗

+434.5%

53,350 → 285,139 downloads

03

vaquill/open-india-law

Open India Law is a dataset of open, structured Indian primary law, together with the scrapers used to build it. It covers every judgment of the Supreme Court of India and all 25 High Courts, decisions of 15 tribunals and regulators, and Central, State and Union Territory legislation down to the individual section.

Hugging Face dataset

Open source ↗

+400.7%

1,200 → 6,008 downloads

04

QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4

QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 is a 4-bit NVFP4 quantized version of Qwen/Qwen3.8-27B, produced with QUASAR, a quantization-aware training method. The NVFP4 weights are trained directly against the frozen BF16 model and exported as standard NVFP4 weights with no custom inference path, covering all 496 transformer linears including attention and gated delta-net. It runs with vLLM and requires an NVIDIA GPU with FP4 support.

Hugging Face model

Open source ↗

+349.8%

5,516 → 24,813 downloads

05

BreezeBlue/Breeze-TTS-2

Breeze TTS 2 is an open-weight text-to-speech model built for real-time interaction. It supports voice cloning from reference audio, creating voices from natural-language descriptions without reference audio, and steering tone, emotion, pace, and delivery of a cloned voice. The model generates English and Chinese speech, supports inline vocal events such as (laugh) and (sigh), and offers low-latency streaming output. Source code is under the Apache License 2.0, while model weights use a separate non-commercial license.

Hugging Face model

Open source ↗

+334.8%

2,236 → 9,722 downloads

06

peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP

Tiel-Coder-35B-A3B-GGUF-MTP is a quantized local language model built for coding work. It is a 35B MoE model distributed as GGUF files at 4-bit quantization, about 23 GB, with an optional MTP draft head enabled in llama.cpp via --spec-type draft-mtp. The build targets agentic coding and multi-turn conversation, and it also supports vision input. The MoE architecture lets it run fast even on smaller GPUs with partial offloading, and it can hold large contexts such as 131k to 262k tokens.

Hugging Face model

Open source ↗

+296.5%

89,522 → 354,918 downloads

07

Specific-Labs/Scaffold-CoT

Scaffold-CoT is a dataset of structured chain-of-thought training data hosted on Hugging Face, containing 3,726,548 examples across 76 JSONL shards. Each row has four top-level fields: metadata, input messages, cot events, and output answers. The cot field records ordered events of type reasoning, tool_call, tool_result, or assistant_note, and index fields record each event's position so multi-turn examples can be reconstructed. The build removed system messages and excluded incomplete, malformed, or instruction-contaminated rows.

Hugging Face dataset

Open source ↗

+277.8%

487 → 1,840 downloads

08

LightwheelAI/EgoPro

EgoPro is a dataset of egocentric video and pose data, forming the 10,000-hour head-and-wrist line of EgoSuite-Open100K. It pairs synchronized head- and wrist-view video with 3D hand pose, and its body subset adds full-body pose. The data is distributed through the LightwheelAI/EgoPro Bucket, with the Git repository serving as the dataset card and access point.

Hugging Face dataset

Open source ↗

+276.7%

9,435 → 35,545 downloads

09

peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF

Tiel-Coder-35B-A3B-GGUF is a quantized GGUF release of a 35B MoE coding language model, distributed in multiple quantization tiers. It is aimed at agentic coding and multi-turn conversation, and inherits a vision tower so it can accept images such as screenshots or stack traces. The release includes a shared BF16 vision projector file and sampling recommendations, and runs with partial GPU offloading on smaller GPUs.

Hugging Face model

Open source ↗

+255.9%

105,974 → 377,203 downloads

10

OBLITERATUS/Ornith-1.5-9B-OBLITERATED

Ornith-1.5-9B-OBLITERATED is a 9B language model based on the Qwen3.5 hybrid architecture, created by removing the safety alignment of Ornith-1.5-9B through an abliteration process so the model responds to most prompts without refusal. It supports a thinking mode and is distributed in bf16 Transformers and GGUF formats. The model card states it is aimed at alignment researchers, red-teamers, and developers who handle safety externally, and notes reduced MMLU and partially degraded function calling.

Hugging Face model

Open source ↗

+233.4%

16,881 → 56,281 downloads

11

mercor/apex-agents

APEX-Agents is a benchmark dataset from Mercor for evaluating whether AI agents can execute long-horizon, cross-application professional services tasks. Its tasks were created by investment banking analysts, management consultants, and corporate lawyers, and require agents to work in realistic environments with files and tools such as docs, spreadsheets, PDFs, email, chat, and calendar. The dataset contains 480 tasks across 33 worlds, spanning banking and consulting job categories.

Hugging Face dataset

Open source ↗

+194%

49,309 → 144,974 downloads

12

unsloth/GLM-5.3-GGUF

unsloth/GLM-5.3-GGUF is a GGUF-format release of the GLM-5.3 language model published by Unsloth. GLM-5.3 uses the same base model as GLM-5.2, with its changes coming from post-training. The model supports controlling its thinking budget through a reasoning_effort parameter that accepts low, high, and max levels, defaulting to max, and it can be run in Unsloth Desktop with toggles for those thinking levels.

Hugging Face model

Open source ↗

+160.7%

41,221 → 107,447 downloads

13

alibaba-pai/MiniMax-H3-Acc-LoRAs

MiniMax-H3-Acc-LoRAs is a set of acceleration LoRA checkpoints for the MiniMax-H3 video generation model. They apply Parallel Decoding Distillation (PDD) to MiniMax-H3, enabling efficient video generation in only a few inference steps. Checkpoints are provided for both FL2VA and Ref2VA variants, including 4-step Turbo and 8-step Acc versions, and are loaded via an apply_pdd_lora function in the provided prediction scripts.

Hugging Face model

Open source ↗

+151.2%

27,009 → 67,839 downloads

14

peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF

Dirk-Qwen3.8-27B-GGUF is a GGUF-quantized release of Qwen3.8-27B, a dense 27B vision-language model, with vision preserved. It ships with the Sharp chat template, which produces terser answers by default at medium reasoning effort, and keeps the MTP (nextn) head for runtimes supporting multi-token-prediction speculative decoding. Quantization tiers come from Unsloth Dynamic 3.0 above 3 bpw and GSQ-RCO below, and effort levels can be set per request.

Hugging Face model

Open source ↗

+150.3%

32,522 → 81,395 downloads

15

Vyber07/cyber-security

Vyber07/cyber-security is a cybersecurity knowledge base dataset intended for AI training. It covers cybersecurity domains including offensive red teaming, bug bounty exploitation, defensive SOC operations, digital forensics, and AI/LLM security. The dataset is 16 GB in size, contains 507 files, and spans more than 30 domains drawn from more than 15 platforms.

Hugging Face dataset

Open source ↗

+148%

1,491 → 3,698 downloads

16

thomsonreuters/Thomson-1.0-Small

Thomson-1.0-Small is an open-weight causal language model from Thomson Reuters, built on the Qwen3.6-35B-A3B Mixture-of-Experts architecture with 35B total parameters and 3B activated, and a native context length of 262,144. It was developed through continual learning with a focus on legal, tax, and journalism work, using continual pre-training on Thomson Reuters proprietary data such as news, contracts, regulatory filings, case law, and statutes.

Hugging Face model

Open source ↗

+118.8%

1,045 → 2,286 downloads

17

faunix/Qwen3.8-27B-Distillation-40K

Qwen3.8-27B-Distillation-40K is a machine learning dataset published on Hugging Face by Faunix. It contains 40,000 reasoning traces distilled from the Qwen3.8-27B model, which the creators generated by running the model on their own infrastructure. The traces cover 4 domains, with prompts sampled from 12 open-source datasets, and each example is a single JSONL line. The dataset is released under the Apache 2.0 license for use in experiments and model fine-tuning.

Hugging Face dataset

Open source ↗

+110.9%

1,658 → 3,496 downloads

18

ornith-ai/Ornith-1.5-35B-A3B

Ornith-1.5-35B-A3B is a mixture-of-experts language model from the Ornith-1.5 family, built through a self-improvement training process. Rather than using a fixed set of human-curated tasks, the model's training loop continuously generates new tasks, finds strategies for solving them, and improves the policy through reinforcement learning. The model activates about 3B parameters per token and is evaluated on coding, reasoning, and agentic benchmarks such as SWE-bench and Terminal-Bench.

Hugging Face model

Open source ↗

+96.1%

172,695 → 338,723 downloads

This dated edition is generated from the reviewed RiseFinder export. Movement figures retain their original source, period and direction.

Get early signals on the websites, apps, repositories and startups gaining momentum. Join the list for Risefinder email briefings.

By subscribing, you ask us to store your email and send Risefinder updates. You can unsubscribe at any time. Privacy.

Want a sample first? Read past daily emails.