and serve with folded in lora layers in bf16 as it skips the quantization aware training step to requant those layers

Thinking Machines (and others) found that RL on low rank LoRA can match full-finetuning RL thinkingmachines.ai/blog/lora/, without Blackwell GPUs it would be easiest to upcast the mxfp4 MoE layers to bf16 to train

Some personal news: I've joined sophont.med to help build the next generation of open medical foundation models.

We've relaunched medarc.ai, our open science research community. Join us if you want to help advance open medical AI.

And we are hiring.

counterpoint: GPT-5 does this, it says it doesn’t know rather than hallucinate, the world hasn’t fallen apart

ModernBERT goes MULTILINGUAL!

One of the most requested models I've seen, @jhuclsp.bsky.social has trained state-of-the-art massively multilingual encoders using the ModernBERT architecture: mmBERT.

Stronger than an existing models at their sizes, while also much faster!

Details in 🧵

ChatGPT has been the best technical search engine since o4-mini.

Thinking Mini still makes for a good faster search if you don’t need the extra reasoning ability.

GPT 5 Thinking (the smartest one) ignored the low quality sources and only cited the high quality and reliable sources.

A modern example: When attempting to trick GPT 5 + search with a question on the health benefits of raw milk, GPT 5 Fast (the less smart one) started out by citing the raw milk institute before eventually concluding there aren’t any benefits and citing high quality sources.

Good LLMs do know and/or can reason about these things. Small, cheap, and often free LLMs are the models which cannot.

Remember the glue on pizza Reddit post that the subpar Google AI cited uncritically? Bing’s then integration of GPT 3.5 recognized the Reddit post as sarcasm.

Writing Speed-of-Light Flash Attention for 5090 in CUDA C++ by Thien Tran

He walkthrough how he learned to implement Flash Attention for 5090 in CUDA C++. The main objective is to learn writing attention in CUDA C++,

Microsoft made a useful LLM copilot tool that could summarize text in spreadsheets. They provided clear instructions about how to use it and not to use it. In response, journalists are now mocking them for doing exactly the right thing and showing how to use and not use the tools.

Reports of AI eating entry level jobs are greatly exaggerated.

My guess is current and near-future LLMs are more likely to increase the demand for programmers, not decrease demand (Jevons Paradox).

There isn't a canonical version, but there are retrieval models from GTE and Nomic which might work for your task.

GTE: huggingface.co/Alibaba-NLP/...
Nomic: huggingface.co/nomic-ai/mod...

For more details, including our simple training method, see Benjamin Clavié's twitter announcement, our model, blog post, and paper.

Twitter: x.com/bclavie/stat...
Model: huggingface.co/answerdotai/...
Blog: www.answer.ai/posts/2025-0...
Paper: arxiv.org/abs/2502.03793

Can all encoders be instruction-tuned? Can we replicate ModernBERT's results with an older model like RoBERTa or peer model like GTE-en-MLM?

No. And it's not close.

When we finetune ModernBERT-Large-Instruct on task specific datasets, the generative MLM head is better or nearly equal to standard classification heads.

After instruction tuning on Flan, ModernBERT-Large-Instruct outperforms similarly sized LLMs on MMLU & MMLU-Pro, and achieves ~90 percent of Llama 3.2 1B's performance with ~65 percent fewer parameters.

It's All in The [MASK]: Simple Instruction-Tuning Enables BERT-like Masked Language Models As Generative Classifiers While encoder-only models such as BERT and ModernBERT are ubiquitous in real-world NLP applications, their conventional reliance on task-specific classification heads can limit their applicability com...

With @bclavie.bsky.social and @ncoop57.bsky.social, we tried to answer two questions:

- Can an instruction-tuned ModernBERT zero-shot tasks using the MLM-head?
- Could we then fine-tune instruction-tuned ModernBERT to complete any task?

Detailed answers: arxiv.org/abs/2502.03793

from transformers import pipeline model_name = "answerdotai/ModernBERT-Large-Instruct" fill_mask = pipeline("fill-mask", model=model_name, tokenizer=model_name) text = """You will be given a question and options. Select the right answer. QUESTION: If (G, .) is a group such that (ab)^-1 = a^-1b^-1, for all a, b in G, then G is a/an CHOICES: - A: commutative semi group - B: abelian group - C: non-abelian group - D: None of these ANSWER: [unused0] [MASK]""" results = fill_mask(text) answer = results[0]["token_str"].strip() print(f"Predicted answer: {answer}") # Answer: B

One of the questions we debated while training ModernBERT was whether a modern trained encoder would unlock zero-shot reasoning using only it's generative head?

Spoilers: the answer is yes.

o3-mini is really good at writing internal documentation - feed it a codebase, get back a detailed explanation of how specific aspects of it work simonwillison.net/2025/Feb/5/o...

The latest open artifacts (#6): Reasoning models, China's lead in open-source, and a growing multimodal space Artifacts log 6 The open LM ecosystem yet again accelerates.

If you want to quickly catch up on all the open modeling things (DeepSeek, ModernBERT, etc.), this was a great overview, by @natolambert.bsky.social.

I somehow got into an argument last week with someone who was insisting that all models are industrial blackboxes... and I wish I'd had this on hand.

Alibaba-NLP/gte-modernbert-base · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

You can find the models on Hugging Face here:

- gte-modernbert-base: huggingface.co/Alibaba-NLP/...
- gte-reranker-modernbert-base: huggingface.co/Alibaba-NLP/...

In addition to being the best retrieval model under 300M params on METB (without extra work), and top 10 for under 1B, here's a fun tidbit from Alibaba's GTE ModernBERT model card:

gte-modernbert-base beats gte-qwen1.5-7b on LoCo long context retrieval with 7B less parameters.

The newest extremely strong embedding model based on ModernBERT-base is out: `cde-small-v2`. Both faster and stronger than its predecessor, this one tops the MTEB leaderboard for its tiny size!

Details in 🧵

ModernBERT-embed-base is awesome because it allows to use ModernBERT-base for various tasks out-of-the-box
But the large variant of ModernBERT is also awesome...
So today, @lightonai.bsky.social is releasing ModernBERT-embed-large, the larger and more capable iteration of ModernBERT-embed!

Finally, a Replacement for BERT: Introducing ModernBERT We’re on a journey to advance and democratize artificial intelligence through open source and open science.

What's ModernBERT? It's a drop-in replacement for existing BERT models, but smarter, faster, and supports longer context.

Check out our announcement post for more details: huggingface.co/blog/modernb...

Transformers v4.48.0: ModernBERT, Aria, TimmWrapper, ColPali, Falcon3, Bamba, VitPose, DinoV2 w/ Registers, Emu3, Cohere v2, TextNet, DiffLlama, PixtralLarge, Moonshine

ModernBERT is officially released on Transformers v4.48.0. You no longer need to install from git to use.

If you are plugging ModernBERT into an existing encoder finetuning pipeline, try increasing the learning rate. We've found that ModernBERT tends to prefer a higher LR than older models.

*Actually, that’s good compared to the 4090’s PCIe 4 without NVLink

The good: 32GB
The bad: $2,000
The Ugly*: PCIe 5 without NVLink

Basically, a frontier model like OpenAI’s O1 is like a Ferrari SF-23. It’s an obvious triumph of engineering, designed to win races, and that’s why we talk about it. But it takes a special pit crew just to change the tires and you can’t buy one for yourself. In contrast, a BERT model is like a Honda Civic. It’s also an engineering triumph, but more subtly, since it is engineered to be affordable, fuel-efficient, reliable, and extremely useful. And that’s why they’re absolutely everywhere.

Via @simonwillison.net's excellent blog, I found this great quote about AI models, from @benjaminwarner.dev et al. www.answer.ai/posts/2024-1...

It seems to me that AI will be most relevant in people's lives because the Honda Civic is ubiquitous, not so much because everyone is driving a Ferrari.