Victoria X Lin

MTS @thinkymachines | Native Multimodal Intelligence

Prev: @AIatMeta @SFResearch • PhD @uwcse

[Victoria X Lin @VictoriaLinML]

Inkling is a 975B-41B(A) MoE model that natively reasons across modalities (text, images and audio). It is intelligent and versatile🌱.

I’ve had so much fun building it alongside an incredible team over the past few months and proud to openly share this work.

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[Mira Murati @miramurati]

Inkling-Small is comparable to Inkling at a quarter the size. Weights are open, fine-tunable on Tinker today. Look forward to seeing what people make with it.

[NVIDIA AI @NVIDIAAI]

Another open-weight release from @thinkymachines 👀 Inkling-Small is here.

With native reasoning over audio and images and variable thinking effort, it's a great choice for fine-tuning, with NVIDIA NeMo on NVIDIA DGX Station.

NVFP4 checkpoint here: https://t.co/oJDeCVGy5p

[Thinking Machines @thinkymachines]

Today, we are releasing Inkling-Small.

Inkling-Small achieves comparable performance to Inkling at a quarter of its size. It features 276B total parameters, 12B active. We are making the full weights available.

https://t.co/BtYNcpkDRA

Fine-tune it on Tinker today, or chat with it in text, image, and audio on Tinker Playground.

[Horace He @cHHillee]

Presenting my grand unified theory of ML researcher impact: Your impact is directly proportional to how much pain you cause to infra.

Fundamentally, you can only inflict pain upon infra if your approach actually works. And the better your approach works the more pain infra is forced to endure.

So, to give some examples:
- MoE's add a ton of data-dependent computation => pain (shazeer++)
- GDN/KDA are the most complex architecture I've been forced to care about and a very annoying matrix inversion => pain (sonta++)
- Muon is much more annoying than Adam and causes annoying restrictions on parallelism => pain (keller/jeremy++)
- RL scaling forced many researchers to care about LLM inference and RL infra as a category => pain (tworek++)

Even papers like Attention Is All You Need have lead to significant pain! Before transformers were invented everyone was running small jobs and I never needed to think about kv-caches or 6D parallelism.

[熊师傅 weight decay 了吗 @bigeagle_xd]

Fun fact: I'm responsible of editing the limitations part, and this final line was written by Zhilin himself.

[Yue Yang @YueYangAI]

Inkling is very good at using tools to solve vision tasks.

In this demo, Inkling relies solely on Python tools to generate segmentation masks, iteratively zooming in, cropping, and refining object boundaries.

Plug Inkling into your multimodal agent workflow!
https://t.co/rWEAnwqb73

[Keiran Paster @keirp1]

The progress has been extremely fast. Most of all, I am happy to be surrounded by good people on a critical mission.

Please hold us to your highest standards and share your feedback as you get to know Inkling!

[Martin Ziqiao Ma @ziqiao_ma]

Meet Inkling, our open-weights model built to reason across words, images, sound, and tasks. I'm too tired to find the words, so I'll let Inkling speak for me.

[Victoria X Lin @VictoriaLinML]

@jonlachman Same here!

[Victoria X Lin @VictoriaLinML]

Inkling is a 975B-41B(A) MoE model that natively reasons across modalities (text, images and audio). It is intelligent and versatile🌱.

I’ve had so much fun building it alongside an incredible team over the past few months and proud to openly share this work.

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[Long Lian @LongTonyLian]

Honored to be in the agentic efforts in Inkling. I felt those sleepless nights were well-spent.

Inkling is just a checkpoint of our progress for the foundation model effort started this January. This is just a preview, and the best is yet to come. ✨

[Saurabh Garg @saurabh_garg67]

Training Inkling was a lot of fun, we hope tinkering with it is too!

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[Lilian Weng @lilianweng]

new post on harness engineering for AI self-improvement: https://t.co/ZYvGfVs61k

It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.

Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.

[aidan @aidanmantine]

today: a humans& blog post from thinky, a thinky blog post from humans&

[John Schulman @johnschulman2]

We started Thinking Machines a year and a half ago with a couple of instincts: that people should have much more ability to customize models and do research on them, and that even as AI becomes more autonomous, there's a lot more to build to make humans and AIs work well together.

A lot has happened since then, especially the massive progress in agents, so we wanted to revisit those instincts in light of everything we've learned, argue about them, and write down what we actually believe now.

This is where we landed after a lot of debate. I'm happy with it!

[Thinking Machines @thinkymachines]

We're building AI that people and organizations can shape and make their own. AI should extend our will and judgment instead of neglecting it; enabling that is the technical challenge we are working to solve.
https://t.co/Bi558y4vqD

[Victoria X Lin @VictoriaLinML]

If you're excited about efficient parallel reasoning 🧠⚡ in LLMs, don't miss the ThreadWeaver paper presentation at #ICML2026.
🗓️Oral: Thu, Jul 9, 2026 • 10:15 AM – 10:30 AM KST
🗓️Poster: Thu, Jul 9, 2026 • 2:30 PM – 4:15 PM KST
📎 https://t.co/M6WoVdYN9f

Feel free to reach out to us if you have any questions (unfortunately I won't be attending in person).

[Long Lian @LongTonyLian]

ThreadWeaver 🧵⚡️ is accepted for an oral presentation at #ICML2026 🎉

ThreadWeaver achieves faster reasoning end-to-end though multiple reasoning agents efficiently working together.

Come to our talk at Thursday, Jul 9, 10 AM and poster at 2:30 PM!

[Victoria X Lin @VictoriaLinML]

Thanks @stevenyfeng @karanps and the class for having me. Hope the discussion is helpful. You could also find the slides at: https://t.co/dgpn0APhGX

[Victoria X Lin @VictoriaLinML]

The video of my Stanford CS25 guest lecture, From Language Models to Native Multimodal Intelligence, is now online.

I discussed how the core ideas behind LLMs has shaped multimodal AI, from architectures to training paradigms and scaling, and where the next challenges may lie. 🧠🌐
🎥: https://t.co/sUVxSEjrVf

[Victoria X Lin @VictoriaLinML]

A personal take ☕: The complexity of the problem space demands models specialized for specific capabilities (e.g. visual reasoning, knowledge processing or embodiment) rather than a single uniform design for everything. Seamless integrating these capabilities would be truly exciting.