←
Horace He3,862 posts
Horace He avatar✓

Horace He ✓

@cHHillee

@thinkymachines Formerly @PyTorch "My learning style is Horace twitter threads" - @typedfemale

◷ Joined February 2010▧ 454 Photos and videos
Horace He avatar
Horace He@cHHillee·10 Sep 2025

Apologies that I haven't written anything since joining Thinking Machines but I hope this blog post on a topic very near and dear to my heart (reproducible floating point numerics in LLM inference) will make up for it!

Post attachment
Jack Cook avatar
⟳ Horace He retweeted
Jack Cook@jackcookjack·Nov 13

Inkling 2 Small is our first model trained partly in NVFP4, with a mixed-precision recipe that enables a great tradeoff between speed and accuracy. Check out our new post to see how we did it, and the lessons we learned along the way!

Horace He avatar
Horace He@cHHillee·Nov 13

When I first talked to @jackcookjack, he thought that although he was having fun working on nvfp4 pretraining in academia, he was skeptical that it would ever actually work for actual large-scale runs. As it turns out, he was wrong :^) This blog post explains how it can be done

Horace He avatar
Horace He@cHHillee·Nov 13

I'll be giving a talk tomorrow at the SGLang Summit at 10am! Oftentimes, I find explanations in ML infra to be unsatisfying. A common failure mode is that they'll explain the *literal* tradeoffs but fail to provide a mental model for how to reason about the choice. This talk won't explain everything about inference (only 15 minutes!), but I'll provide two mental models I've found useful when reasoning about inference, and I hope you find them useful too :)

Post attachment
Horace He avatar
Horace He@cHHillee·Nov 12

Inkling 2 small weights are out! Go have fun with it yourself, or continue using Tinker :^)

Jasper Dekoninck avatar
⟳ Horace He retweeted
Jasper Dekoninck@j_dekoninck·Nov 12

We just added Inkling 2 Small to MathArena! Strong performance for an open model, especially on BrokenArXiv, where it takes away the top spot among open models.

Post attachment
Horace He avatar
Horace He@cHHillee·Nov 10

Note, please also explain (at least briefly) what you want to do LOL. This is not meant be an airdrop hahaha.

Horace He avatar
Horace He@cHHillee·Nov 10

Ahead of our open weights release, we'd like to encourage folks to try out the model for cool things, whether that's evals, building a cool demo, or posttraining it! If you have some good ideas, comment below or DM your email and we can give you some tinker credits to try it :)

Horace He avatar
Horace He@cHHillee·Nov 9

(Also, for what purpose do you want such long context lengths mainly?)

Horace He avatar
Horace He@cHHillee·Nov 9

How valuable would training contexts longer than 256k be on Tinker? And if you want longer than 256k, how long would you want them to be?

rahul avatar
⟳ Horace He retweeted
rahul@0interestrates·Nov 7

Inkling 2 Small pushes the cost-performance pareto frontier for OSS models on DeepFinance bench On long horizon finance workflows, Inkling 2 Small scores higher than GLM 5.3 for 50% of the cost, and nearly the same score as Qwen 3.8 max for 11% of the cost

Post attachmentPost attachment
Surya avatar
⟳ Horace He retweeted
Surya@kickingkeys·Nov 6

Inkling 2 Small can paint 👩‍🎨!! (Video + Audio made in code) Comparison below between GLM 5.3 Flash, DeepSeek V4.1 Flash and Qwen 3.8 Flash on the same prompts. ⬇️

0:30
0:30
ARC Prize avatar
⟳ Horace He retweeted
ARC Prize@arcprize·Nov 6

Inkling 2 Small from @thinkymachines on ARC-AGI (Verified): - ARC-AGI-2: 81.7%, $0.27/task - ARC-AGI-1: 96.5%, $0.08/task It outperforms every open-weight model we've evaluated on both ARC-AGI-1 & ARC-AGI-2. Thinking Machines plans to release its full weights next week.

Post attachment
Le Zhuo avatar
⟳ Horace He retweeted
Le Zhuo@zhuole1025·Nov 6

Inkling 2 is a great model especially for multimodal tasks! just tested it on internal benchmarks for image generation & editing judging, and it's better than kimi and muse. looking forward to fine-tuning Inkling on Tinker.

Horace He avatar
Horace He@cHHillee·Nov 5

Our infra has really improved quite a bit since the last time we released a model - this one took just under 3 days to pretrain! Also, it's quite cheap on Tinker (0.30/1.00/1.50 for in/out/train)! Hope you guys find this model great for customization purposes :)