Lightnews — Scholar-powered news

Adina Yakup @adinayakup.bsky.social · Sep 9

Baidu just released a reasoning model 🔥 ERNIE-4.5-21B-A3B-Thinking

huggingface.co/baidu/ERNIE-...

✨ Small MoE - Apache 2.0
✨ 128K context length for deep reasoning
✨ Efficient tool usage capabilities

baidu/ERNIE-4.5-21B-A3B-Thinking · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

4 16

Adina Yakup @adinayakup.bsky.social · Sep 9

Baidu just released a thinking model 🔥 ERNIE-4.5-21B-A3B-Thinking

huggingface.co/baidu/ERNIE-...

✨ Small MoE - Apache 2.0
✨ 128K context length for deep reasoning
✨ Efficient tool usage capabilities

baidu/ERNIE-4.5-21B-A3B-Thinking · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

3

Adina Yakup @adinayakup.bsky.social · Sep 6

MiniCPM4.1🔥New edge-side LLM built for efficiency + reasoning from OpenBMB

huggingface.co/openbmb/Mini...

✨ 8B - Apache 2.0
✨ Hybrid reasoning model: deep reasoning +fast inference.
✨5x faster on edge chips, 90% smaller (BitCPM)
✨Trained on UltraClean + UltraChat v2 data

openbmb/MiniCPM4.1-8B · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

2 5

Adina Yakup @adinayakup.bsky.social · Sep 5

Inverse IFEval 🔥New benchmark from Bytedance & MAP

huggingface.co/datasets/m-a...
huggingface.co/papers/2509....

Testing LLMs on their ability to override biases & follow adversarial instructions.
✨ 8 challenge types
✨ 1,012 CN/EN Qs across 23 domains
✨ Human-in-the-loop + LLM-as-a-Judge

m-a-p/Inverse_IFEval · Datasets at Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Adina Yakup @adinayakup.bsky.social · Sep 5

Added the Kwai team from Kuaishou in the heatmap
huggingface.co/spaces/zh-ai...

Chinese Open Source Heatmap - a Hugging Face Space by zh-ai-community

This application shows a heatmap of the most active AI labs and organizations on Hugging Face, highlighting their recent model releases. Users can see which labs are most active and their contribut...

huggingface.co

1

Adina Yakup @adinayakup.bsky.social · Sep 5

Klear-46B-A2.5🔥 a sparse MoE LLM developed by the Kwai-Klear Team at Kuaishou

huggingface.co/collections/...

✨ 46B total / 2.5B active - Apache2.0
✨ Dense-level performance at lower cost
✨ Trained on 22T tokens with progressive curriculum
✨ 64K context length

Klear1.0 - a Kwai-Klear Collection

Klear1.0

huggingface.co

2 2 5

Adina Yakup @adinayakup.bsky.social · Sep 5

Latest update from Moonshot AI

Kimi K2 >>> Kimi K2-Instruct-0905🔥

huggingface.co/moonshotai/K...

✨ 32B activated / 1T total parameters
✨ Enhanced agentic coding intelligence
✨ Better frontend coding experience
✨ 256K context window for long horizon tasks

moonshotai/Kimi-K2-Instruct-0905 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

2 6

Adina Yakup @adinayakup.bsky.social · Sep 1

✨ Supports 33 languages, including 5 ethnic minority languages in China 👀
✨ Including a translation ensemble model: Chimera-7B
✨ Full pipeline: pretrain > CPT > SFT > enhancement > ensemble refinement > SOTA performance at similar scale

1 5

Adina Yakup @adinayakup.bsky.social · Sep 1

Hunyuan-MT-7B 🔥 open translation model released by Tencent

huggingface.co/collections/...

Hunyuan-MT - a tencent Collection

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

2 1 8

Adina Yakup @adinayakup.bsky.social · Sep 1

✨ 560B total / ~27B active MoE — MIT license
✨ 128k context length + advanced reasoning
✨ ScMoE design: 100+ TPS inference
✨ Stable large-scale training + strong agentic performance

Adina Yakup @adinayakup.bsky.social · Sep 1

From food delivery to frontier AI 🚀 Meituan, the leading lifestyle platform just dropped its first open SoTA LLM: LongCat-Flash 🔥

huggingface.co/meituan-long...

meituan-longcat/LongCat-Flash-Chat · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 1 4

Adina Yakup @adinayakup.bsky.social · Aug 29

✨ Large-scale triplet dataset (content, style, stylized)
✨ Disentangled learning: style alignment + content preservation
✨ Style Reward Learning (SRL) for higher fidelity
✨ USO-Bench: 1st benchmark for style & subject jointly
✨ SOTA results on subject consistency & style similarity

Adina Yakup @adinayakup.bsky.social · Aug 29

USO 🎨 Unified customization model released by Bytedance research

Demo
huggingface.co/spaces/byted...
Model
huggingface.co/bytedance-re...
Paper
huggingface.co/papers/2508....

USO - a Hugging Face Space by bytedance-research

Create custom images by combining different styles and subjects. Upload content and style images, or use text prompts to generate photorealistic portraits or styled images.

huggingface.co

1 1

Adina Yakup @adinayakup.bsky.social · Aug 29

✨ Direct raw audio: text & speech ,no ASR+LLM+TTS pipeline
✨ High-IQ reasoning: RL + CoT for paralinguistic cues
✨ Multimodal RAG + tool calling
✨ Emotion, timbre, dialect & style control
✨ SOTA on ASR, paralinguistic, speech dialog

1

Adina Yakup @adinayakup.bsky.social · Aug 29

Step-Audio 2🔥 New end to end multimodal LLM for audio & speech, released by StepFun

huggingface.co/collections/...

Step-Audio 2 - a stepfun-ai Collection

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 2 6

Adina Yakup @adinayakup.bsky.social · Aug 26

>Applications: AI-as-a-service, test bases, new standards
>Open-source: support communities, encourage contributions (incl. university credits & recognition), foster new application approaches, and build globally impactful ecosystems 👀
>Talent, policy & safety frameworks: secure sustainable growth

1

Adina Yakup @adinayakup.bsky.social · Aug 26

✨Highlights:
>Models: advance theory, efficient training/inference, evaluation system
>Data: high-quality datasets, IP/copyright reform, new incentives
>Compute: boost chips & clusters, improve national network, promote cloud standardization, and ensure inclusive, efficient, green, secure supply.

1 3

Adina Yakup @adinayakup.bsky.social · Aug 26

🇨🇳 China’s State Council just released its “AI+” Action Plan (2025)

huggingface.co/spaces/zh-ai...

✨Goal: By 2035, AI will deeply empower all sectors, reshape productivity & society

✨Focus on 6 pillars:
>Science & Tech
>Industry
>Consumption
>Public welfare
>Governance
>Global cooperation

China AI policy research 🤗 - a Hugging Face Space by zh-ai-community

Browse and filter through key AI policy documents from China, covering various topics like domestic development, international cooperation, safety, data management, industry standards, and ethics. ...

huggingface.co

1 11

Adina Yakup @adinayakup.bsky.social · Aug 26

✨ SOTA vision language capability
✨ 96× video token compression > high-FPS & long video reasoning
✨ Switchable fast vs deep thinking modes
✨ Strong OCR, document parsing, supports 30+ languages

1

Adina Yakup @adinayakup.bsky.social · Aug 26

MiniCPM-V 4.5 🚀 New MLLM for image, multi-image & video understanding, running even on your phone, released by OpenBMB

huggingface.co/openbmb/Mini...

openbmb/MiniCPM-V-4_5 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 2 5

Adina Yakup @adinayakup.bsky.social · Aug 26

InternVL3.5 🔥 New family of multimodal model by Shanghai AI lab @opengvlab

huggingface.co/collections/...

✨ 1B · 2B · 4B · 8B · 14B · 38B ｜ MoE → 20B-A4B · 30B-A3B · 241B-A28B 📄Apache 2.0
✨ +16% reasoning performance, 4.05× speedup vs InternVL3

InternVL3.5 - a OpenGVLab Collection

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 3

Adina Yakup @adinayakup.bsky.social · Aug 21

Intern-S1-mini 🔥 lightweight open multimodal reasoning model by Shanghai AI Lab.

huggingface.co/internlm/Int...

✨ Efficient 8B LLM + 0.3B vision encoder
✨ Apache 2.0
✨ 5T multimodal pretraining, 50%+ in scientific domains
✨ Dynamic tokenizer for molecules & protein sequences

internlm/Intern-S1-mini · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1 4

Adina Yakup @adinayakup.bsky.social · Aug 20

Seed-OSS 🔥 The latest open LLM from Bytedance Seed team

huggingface.co/collections/...

✨ 36B - Base & Instruct
✨ Apache 2.0
✨ Native 512K long context
✨ Strong reasoning & agentic intelligence
✨ 2 Base versions: with & without synthetic data

Seed-OSS - a ByteDance-Seed Collection

Seed-OSS Open-Source Models

huggingface.co

3

Adina Yakup @adinayakup.bsky.social · Aug 19

✨DeepSeek V3.1 just dropped on @hf.co

huggingface.co/collections/...

DeepSeek-V3.1 - a deepseek-ai Collection

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

2 14

Adina Yakup @adinayakup.bsky.social · Aug 18

✨ Apache 2.0
✨ Semantic + Appearance Editing: rotate, restyle, add/remove
✨ Precise Text Editing → edit CN/EN text, keep style
huggingface.co/spaces/Qwen/...

Qwen Image Edit - a Hugging Face Space by Qwen

This application allows you to edit images by providing instructions. You upload an image and describe the changes you want, and the app generates the edited image for you. You can specify addition...

huggingface.co

1 2