TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification
Paper • 2604.14531 • Published • 6
AI & ML interests
Hugging Face Inference Endpoints Images repository allows AI Builders to collaborate and engage creating awesome inference deployments
Recent Activity
View all activity
AdinaY
submitted a
paper to Daily Papers 2 days ago
qgallouedec
posted an update 2 days ago
Post
1599
TRL v1.2 introduces the SSDTrainer 🚀
Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.
The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.
v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of
Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0
Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.
The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.
from trl.experimental.ssd import SSDConfig, SSDTrainer
trainer = SSDTrainer(
model="Qwen/Qwen3-4B-Instruct",
args=SSDConfig(temperature=0.6, top_k=20, top_p=0.95),
train_dataset=dataset,
)
trainer.train()v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of
use_transformers_paged, and key fixes for VLM response parsing.Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0
qgallouedec
posted an update 19 days ago
Post
2305
TRL v1.0 is out!
Hugging Face's TRL library is downloaded 3 million times a month. Over 130k models trained with it are public on the Hub, and major projects like @unsloth and @axolotl-ai-co build directly on top of it. v1.0 is the moment we acknowledged that responsibility explicitly, with a real stability contract.
The field hasn't settled. Building stable software in a domain that keeps invalidating its own assumptions is the actual problem we're solving. The answer is a design that can absorb the next shift without breaking what people rely on.
What's in v1.0:
Deep Hugging Face integration, low infrastructure burden
What's next: asynchronous GRPO, better scaling support, and making training legible enough that agents can inspect and steer it.
Read more: hf.co/blog/trl-v1
Hugging Face's TRL library is downloaded 3 million times a month. Over 130k models trained with it are public on the Hub, and major projects like @unsloth and @axolotl-ai-co build directly on top of it. v1.0 is the moment we acknowledged that responsibility explicitly, with a real stability contract.
The field hasn't settled. Building stable software in a domain that keeps invalidating its own assumptions is the actual problem we're solving. The answer is a design that can absorb the next shift without breaking what people rely on.
What's in v1.0:
Deep Hugging Face integration, low infrastructure burden
What's next: asynchronous GRPO, better scaling support, and making training legible enough that agents can inspect and steer it.
pip install --upgrade trlRead more: hf.co/blog/trl-v1
AdinaY
submitted a
paper to Daily Papers 19 days ago
AdinaY
submitted 2
papers to Daily Papers about 1 month ago
alvarobartt
posted an update about 1 month ago
Post
3603
Learn how to deploy Microsoft Research VibeVoice ASR on Microsoft Azure Foundry with Hugging Face to generate rich audio transcriptions with Who, When, and What! 💥
> 🕒 60-minute single-pass processing, no chunking or stitching
> 👤 Customized hotwords to guide recognition on domain-specific content
> 📝 Rich transcription: joint ASR + diarization + timestamping in one pass
> 🌍 50+ languages with automatic detection and code-switching support
> 🤗 Deployed on Microsoft Foundry via an OpenAI-compatible Chat Completions API
https://huggingface.co/docs/microsoft-azure/foundry/examples/deploy-vibevoice-asr
> 🕒 60-minute single-pass processing, no chunking or stitching
> 👤 Customized hotwords to guide recognition on domain-specific content
> 📝 Rich transcription: joint ASR + diarization + timestamping in one pass
> 🌍 50+ languages with automatic detection and code-switching support
> 🤗 Deployed on Microsoft Foundry via an OpenAI-compatible Chat Completions API
https://huggingface.co/docs/microsoft-azure/foundry/examples/deploy-vibevoice-asr
qgallouedec
posted an update about 2 months ago
Post
3007
@CohereLabs just released 🌿 Tiny Aya: a fully open-source 3B parameter model that speaks 70+ languages 🌍! But there’s a catch:
Tiny Aya is just a language model. It doesn’t support tool calling, the key capability that turns frontier models into powerful *agents*.
So the real question is:
How hard is it to turn Tiny Aya into an agent?
Turns out… it’s simple, thanks to Hugging Face TRL.
We’re sharing a hands-on example showing how to train Tiny Aya to turn it into a tool-calling agent using TRL, unlocking what could become the first *massively multilingual open agent*.
Small model. Global reach. Agent capabilities.
👉 https://github.com/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb
Tiny Aya is just a language model. It doesn’t support tool calling, the key capability that turns frontier models into powerful *agents*.
So the real question is:
How hard is it to turn Tiny Aya into an agent?
Turns out… it’s simple, thanks to Hugging Face TRL.
We’re sharing a hands-on example showing how to train Tiny Aya to turn it into a tool-calling agent using TRL, unlocking what could become the first *massively multilingual open agent*.
Small model. Global reach. Agent capabilities.
👉 https://github.com/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb
- 1 reply
Post
3630
MiniMax M2.5 is now available on the hub 🚀
MiniMaxAI/MiniMax-M2.5
✨ 229B - Modified MIT license
✨37% faster than M2.1
✨ ~$1/hour at 100 TPS
MiniMaxAI/MiniMax-M2.5
✨ 229B - Modified MIT license
✨37% faster than M2.1
✨ ~$1/hour at 100 TPS
- 2 replies
Post
739
RynnBrain 🤖 a physics aware embodied brain for robots from Alibaba DAMO
https://huggingface.co/collections/Alibaba-DAMO-Academy/rynnbrain
✨ 2B/8B/30B (3B active)
✨ Apache 2.0
✨ Understands egocentric scenes with strong spatial awareness
✨ Tracks objects and motion over time
https://huggingface.co/collections/Alibaba-DAMO-Academy/rynnbrain
✨ 2B/8B/30B (3B active)
✨ Apache 2.0
✨ Understands egocentric scenes with strong spatial awareness
✨ Tracks objects and motion over time
- 2 replies
Post
4119
Game on 🎮🚀
While Seedance 2.0’s videos are all over the timeline, DeepSeek quietly pushed a new model update in its app.
GLM-5 from Z.ai adds more momentum.
Ming-flash-omni from Ant Group , MiniCPM-SALA from OpenBMB
, and the upcoming MiniMax M2.5 keep the heat on 🔥
Spring Festival is around the corner,
no one’s sleeping!
✨ More releases coming, stay tuned
https://huggingface.co/collections/zh-ai-community/2026-february-china-open-source-highlights
While Seedance 2.0’s videos are all over the timeline, DeepSeek quietly pushed a new model update in its app.
GLM-5 from Z.ai adds more momentum.
Ming-flash-omni from Ant Group , MiniCPM-SALA from OpenBMB
, and the upcoming MiniMax M2.5 keep the heat on 🔥
Spring Festival is around the corner,
no one’s sleeping!
✨ More releases coming, stay tuned
https://huggingface.co/collections/zh-ai-community/2026-february-china-open-source-highlights
Post
3949
Ming-flash-omni 2.0 🚀 New open omni-MLLM released by Ant Group
inclusionAI/Ming-flash-omni-2.0
✨ MIT license
✨ MoE - 100B/6B active
✨ Zero-shot voice cloning + controllable audio
✨ Fine-grained visual knowledge grounding
inclusionAI/Ming-flash-omni-2.0
✨ MIT license
✨ MoE - 100B/6B active
✨ Zero-shot voice cloning + controllable audio
✨ Fine-grained visual knowledge grounding
- 2 replies
Post
801
LLaDA 2.1 is out 🔥 A new series of MoE diffusion language model released by AntGroup
inclusionAI/LLaDA2.1-mini
inclusionAI/LLaDA2.1-flash
✨LLaDA2.1-mini: 16B - Apache2.0
✨LLaDA2.1-flash: 100B - Apache2.0
✨Both delivers editable generation, RL-trained diffusion reasoning and fast inference
inclusionAI/LLaDA2.1-mini
inclusionAI/LLaDA2.1-flash
✨LLaDA2.1-mini: 16B - Apache2.0
✨LLaDA2.1-flash: 100B - Apache2.0
✨Both delivers editable generation, RL-trained diffusion reasoning and fast inference
- 2 replies
Post
2619
AI for science is moving fast🚀
Intern-S1-Pro 🔬 a MoE multimodal scientific reasoning model from Shanghai AI Lab
internlm/Intern-S1-Pro
✨ 1T total / 22B active
✨ Apache 2.0
✨ SoTA scientific reasoning performance
✨ FoPE enables scalable modeling of long physical time series (10⁰–10⁶)
Intern-S1-Pro 🔬 a MoE multimodal scientific reasoning model from Shanghai AI Lab
internlm/Intern-S1-Pro
✨ 1T total / 22B active
✨ Apache 2.0
✨ SoTA scientific reasoning performance
✨ FoPE enables scalable modeling of long physical time series (10⁰–10⁶)
- 2 replies
Post
1408
✨ China’s open source AI ecosystem has entered a new phase
https://huggingface.co/blog/huggingface/one-year-since-the-deepseek-moment-blog-3
One year after the “DeepSeek Moment,” open source has become the default. Models, research, infrastructure, and deployment are increasingly shared to support large-scale, system-level integration.
This final blog examines how leading Chinese AI organizations are evolving ,and what this implies for the future of open source.
https://huggingface.co/blog/huggingface/one-year-since-the-deepseek-moment-blog-3
One year after the “DeepSeek Moment,” open source has become the default. Models, research, infrastructure, and deployment are increasingly shared to support large-scale, system-level integration.
This final blog examines how leading Chinese AI organizations are evolving ,and what this implies for the future of open source.
Post
415
GLM just entered the OCR field🔥
zai-org/GLM-OCR
✨ 0.9B
✨ MIT licensed
✨ Multimodal GLM-V architecture
✨ #1 on OmniDocBench v1.5 (94.62)
zai-org/GLM-OCR
✨ 0.9B
✨ MIT licensed
✨ Multimodal GLM-V architecture
✨ #1 on OmniDocBench v1.5 (94.62)
Post
1625
Step 3.5 Flash 🔥 new foundation model from StepFun ai
https://huggingface.co/collections/stepfun-ai/step-35-flash
✨ Sparse MoE：196B/11B active
✨ Supports up to 256K context
✨ Multi-token prediction for fast decoding (100–300 tok/s)
✨ Runs locally on consumer hardware
https://huggingface.co/collections/stepfun-ai/step-35-flash
✨ Sparse MoE：196B/11B active
✨ Supports up to 256K context
✨ Multi-token prediction for fast decoding (100–300 tok/s)
✨ Runs locally on consumer hardware
Post
1141
What a week 🤯
Following DeepSeek, Kimi, Qwen, Baidu, and Ant Group, Unitree Robotics
has now released a VLA model on the hub too!
unitreerobotics/UnifoLM-VLA-Base
Following DeepSeek, Kimi, Qwen, Baidu, and Ant Group, Unitree Robotics
has now released a VLA model on the hub too!
unitreerobotics/UnifoLM-VLA-Base
alvarobartt
posted an update 3 months ago
Post
3221
💥
💡 Alternatively, you can also set the
hf-mem v0.4.1 now also estimates KV cache memory requirements for any context length and batch size with the --experimental flag!uvx hf-mem --model-id ... --experimental will automatically pull the required information from the Hugging Face Hub to include the KV cache estimation, when applicable.💡 Alternatively, you can also set the
--max-model-len, --batch-size and --kv-cache-dtype arguments (à la vLLM) manually if preferred. - 1 reply
Post
323
LongCat-Flash-Lite🔥 a non-thinking MoE model released by Meituan LongCat team.
meituan-longcat/LongCat-Flash-Lite
✨ Total 68.5B / 3B active - MIT license
✨ 256k context
✨ Faster inference with N-gram embeddings
meituan-longcat/LongCat-Flash-Lite
✨ Total 68.5B / 3B active - MIT license
✨ 256k context
✨ Faster inference with N-gram embeddings