Hey twitter!

I'm releasing the LLM Evaluation Guidebook v2!
Updated, nicer to read, interactive graphics, etc!
huggingface.co/spaces/OpenEva…

After this, I'm off: I'm taking a sabbatical to go hike with my dogs :D
(back @huggingface in Dec *2026*)

See you all next year!

We need to be dog maxxing our models

There is 0% chance of doom if the chain of thought is

user: hello
model: <think> Oh my god it's a human. I love humans. I will say hi to him, maybe he will give me a problem to chase. I love chasing problems. I will chase the problem for him then he will love me and I will be his best friend. Must help human </think>
model: hello! what can I help you with today?

👀Introducing a brand new @yupp_ai SVG leaderboard ranking frontier models on the generation of coherent and visually appealing SVGs!

Gemini 3 Pro by @GoogleDeepMind takes the crown as the most powerful model! 👏

We’re also releasing a public SVG dataset. Details in🧵

Either you crack general intelligence -- the ability to efficiently acquire arbitrary skills on your own -- or you don't have AGI.

A big pile of task-specific skills memorized from handcrafted/generated environments isn't AGI, not matter how big.

cc @maximelabonne since you wanted an update :P

The guide is very beginner friendly, as we go from the basics of tokenization/inference to the nits and tricks of running eval properly, so it's compatible with all levels.

Should contain most of what we wrote about evals at HF in a single unified place, with updates ofc :)

If you see improvements, I'd love to hear them (within the next 2 days) :)

Many thanks to @thibaudfrere for his help on the banner and @gui_penedo for his proofreading!

If you've got eval needs, your new PoC is @nathanhabib1011 (with a focus on lighteval)!

as a researcher, it makes no sense to compare reasoning vs non reasoning models on benches like the ones in Artificial Analysis without normalizing somehow by cost or output tokens. non reasoning models (base/instruct) are important for the open ecosystem since research teams and companies will use them to do RL or other things (like synthetic generation) for specific verticals (think cursor/windsurf)

as a user, i get that you don’t care whether the model is reasoning or not, you judge speed, cost, and accuracy (and memory if you want to deploy your model locally)

the only advantage of non reasoning models would be speed/cost because they generate fewer tokens BUT speed and cost also depend on other thing like infra
-> for speed see how fast some models get on groq or cerebras
-> for cost model like deepseek are so cheap that there is very few use case where you'd want to use non reasoning model anyway

Mistral has delivered super capable small models but no one is talking about it so here I go

Transformers v5's first release candidate is out 🔥 The biggest release of my life.

It's been five years since the last major (v4). From 20 architectures to 400, 20k daily downloads to 3 million.

The release is huge, w/ tokenization (no slow tokenizers!), modeling &amp; processing.

stop looking at HLE (with tools), most of these mean "has web access"

the answers to HLE are easily accessible in ungated mirrors (and prob a dozen other places). the only question is why those agents don't score 100%

Okay, but, wait, what reasoning traces should I train on?

Excited to share our latest research paper together with @nvidia:
Learning to Reason: Training LLMs with GPT-OSS or DeepSeek R1
Reasoning Traces

arxiv.org/abs/2511.19333 🧵

China just passed the U.S. in open model downloads for the first time 👀

New data from Economies of Open Intelligence led by @huggingface policy team & community collaborators, presents some notable observations:

✨ Developer adoption
In 2025, Chinese model developers saw highthe first time, driven by the rapid rise of @deepseek_ai and @Alibaba_Qwen.

✨ The “Sino-Multimodal Period”(late 2024–present)
China’s share of downloads reached 17.1%, surpassing the U.S., with DeepSeek + Qwen accounting for 14% of recent activity.

This period also brings larger, more quantized, and expanding multimodal models such as Wan2.1.

✨ Organizational patterns
China’s open model development is more industry-driven (similar to the U.S.), while the EU has more university, nonprofit, and community-led contributors.

fyi - this analysis based on 851k models, 200+ attributes, and 2.2B downloads.

Introducing "The Eiffel Tower Llama"!🗼

Remember Golden Gate Claude? Unfortunately Anthropic's viral demo was shut down after 24h, and key technical details remained hidden.

So we recreated it, uncovering key insights on steering LLMs using SAEs⚒️

Full blog post + live demo 👇

Non-natural image gen and editing are difficult tasks. We tested the state of the art at the time — including Nano Banana 1.0 & GPT-image — all performed quite poorly on StructBench.

Nano Banana 2 (NB2) just dropped, and its improvements strongly validate a direction we studied in Ses 90+ on our image generation tasks—by far the best we’ve seen 🔥

A few months before the release of Nano Banana 2, we introduced StructBench — a benchmark for evaluating models on non-natural images like diagrams, math figures, charts, and documents.

Our motivation was simple: today’s image models are overly optimized for aesthetics, but struggle with factuality + structural reasoning. If we want truly unified multimodal models, the training mix needs non-natural data too.

But NB2 still isn’t perfect: we still find failure cases where it misinterprets instructions or misses structural details.

Excited to see the field moving toward models that reason as well as they render.

Below, we provide some more analysis along with cool results!

@GeminiApp @GoogleDeepMind

"Professors definitely deserve to have their names on the papers."

I think this take is completely wrong. Financial support does not warrant co-authorship.

Bob Gallager (a legendary information theorist who retired from MIT) did not co-author any papers with many of his not believe that he made an intellectual contribution that warranted co-authorship.

The screenshot is from Erdal Arıkan's PhD thesis work that was published in IEEE Trans. Information Theory. Both Erdal and Bob have been honored with the Shannon Award (highest honor in information theory) and they have not co-authored any papers.

What happened to adding error bars to evals?

Know anyone who need some help to get started with ML, open source and 🤗?

We partnered with @TechToTheRescue, a tech for good incubator, &amp; answered all AI questions their non profits had to create an FAQ!

github.com/huggingface/faq

Come add your Q/As, it's collaborative! 🔥

incredibly detailed technical blog just dropped on the anatomy of BoltzGen 🧬

made for ML people, but covering everything from molecular representations to diffusion-based generation of protein binders

+ crazy good interactive visuals

👏👏 @ludocomito