christina

yap gear @mit reasoned @xai

[christina @luoluo_ai]

Grok 4 is live: https://t.co/78oPd90REk

[Daniel @nearlydaniel]

War Room squad locked in

[christina @luoluo_ai]

@JunyaoShi @sundayrobotics congrats!!

[christina @luoluo_ai]

@timshi_ai yet the only way to commoditize compute is to enable e2e chip design - from architecture to silicon

[christina @luoluo_ai]

@tararezaeikh @mirendil congrats my day1

[Tara Rezaei @tararezaeikh]

Today, we are formally announcing @mirendil. Mirendil exists to be a straight shot at solving bottlenecks to step-change acceleration across all areas of science and technology. If you are excited about this mission, feel free to reach out.

[Emily Han @emilyhanyf]

we are launching Auto Endpoints ! here are some features supported out of the box, only on @modal:

> sub-5ms high-perf proxy with auth support
> inference-level metrics built into your Modal dashboard (ttft, itl, accept length, etc)
> spec dec (dflash!) and optimized hardware based on the model you're running
> perf preview before deployment, with benchmarks targeting your use case (real-time gen, agentic multi-turn)
> and more !

this is just the beginning -- more autoinference and optimizations are coming soon. time to own your inference !

p.s. every Auto Endpoint user gets $100 in credits. give it a try and let us know what you think !

[christina @luoluo_ai]

@richardczl @ying11231 the goat 🐐

[Radostin Cholakov @radi_cho]

Excited to release cho-embedding-0.8b - a Pareto frontier 0.8B vision-language embedder competitive with some of the most widely-used models up to 4x larger.

Built on Qwen3.5-0.8B and post-trained with multi-stage contrastive + distillation objectives, the model achieves 60.7 on MMEB-Image, the best score among <1B models.

• General-purpose embeddings for text, images, video frames, documents, and document screenshots.
• State-of-the-art performance on MMEB-Image among academic models at this scale (some of which highly specialized to the benchmark).
• Drop-in integration for existing Qwen3-VL-Embedding-2/8B pipelines with significantly lower compute and serving cost.
• Fully open under Apache 2.0 for commercial use.

Download model: https://t.co/iP1sFsppMz

PS. The model name is an inside joke and also an acronym for the "Contrastive Hard-negatives Objective" used. 6720 NVIDIA H100 GPU hours were spent on mining, training, and ablations. Feedback is welcome:)

[Towaki Takikawa / 瀧川永遠希 @yongyuanxi]

Ultrasound tomography…. but with piezoelectric MEMS semiconductors instead of crystals. Presumably imaging algorithms / hardware systems by Midjourney, device design by Butterfly Networks, manufactured by TSMC… super cool!

[christina @luoluo_ai]

@dustinvtran congrats dustin!

[christina @luoluo_ai]

@PascualMerita wow i like this animation

[christina @luoluo_ai]

If you saw where it leads, would you choose the harder path?

[christina @luoluo_ai]

@ryanbai so tough ryan

[christina @luoluo_ai]

@Jackyhuang @hardlaunching @BainCapVC @tomo congrats jacky on finding a lifelong mission 💛

[Elon Musk @elonmusk]

@AJamesMcCarthy Read books, talk to people &amp; iterate rapidly with hardware &amp; software

[christina @luoluo_ai]

@op7418 miss

[christina @luoluo_ai]

love this thesis!! humans are beautiful

[carlo agostinelli @carloagostinel2]

Been fun working on this over the past few weeks with the team.

We wake up every day looking for people who are still unknown to the world but won’t be forever.

The founders we back today will one day have biographies of their own.

@novaholdings

[christina @luoluo_ai]

@TianweiY Amazing!! will stay tune (hopefully a tech report?)

[christina @luoluo_ai]

"With that, we reframed multimodal generation as structured text/code generation"

Text is ambiguous but code is not. Would love to see more results in having LLM natively think like its coding.

[Tianwei Yin @TianweiY]

1/ Our new @reve image model is now #2 on the @arena text-to-image leaderboard — behind only GPT Image 2, ahead of Nano Banana Pro, Microsoft, xAI and everyone else.

And it's a 125 point jump over Reve 1.5 from just 3 months ago.

The research story behind it 🧵👇

[Kasey Zhang @_WEEXIAO]

https://t.co/GI1Q5J0rd5

[John Schulman @johnschulman2]

Glad to see this -- renderers are a foundational component of the LLM stack. Renderers map between tokens and messages, which are invariant to tokenizer and formatting details. Most APIs, datasets, and RL environments are defined in terms of messages.
Getting the details wrong leads to train-test mismatches, caching inefficiencies, and prompt injection vulnerabilities. We included a renderers module in Tinker Cookbook, but it makes sense as a standalone library.