Radical Numerics (@RadicalNumerics) / X

Radical Numerics

10 posts

Radical Numerics

@RadicalNumerics

Systems, scaling and architecture for next-gen scientific world models.

San Francisco & Tokyo

Joined May 2025

Radical Numerics reposted
Michael Poli
@MichaelPoli6
Mar 18
We're growing rapidly at @RadicalNumerics and scaling our core teams. Join us in building the next generation of scientific world models. We're hiring across a few roles, each with significant ownership and cross-functional scope: - Member of Technical Staff, Post-Training -
33K
Radical Numerics
@RadicalNumerics
Jan 12
Scaling scientific world models requires co-designing architectures, training objectives, and numerics. Today, we share the first posts in our series on low-precision pretraining, starting with NVIDIA's NVFP4 recipe for stable 4-bit training. Part 1: radicalnumerics.ai/blog/nvfp4-par… Part
68K
Radical Numerics reposted
Garyk Brixi
@garykbrixi
Oct 14, 2025
Excited to share Phalanx, our new layer for sequence modeling! Each block communicates with its neighbor, like the shield cover of a neighboring hoplite. Phalanx can replace sliding window attention and trains faster than optimized baselines while maintaining quality.
Radical Numerics
@RadicalNumerics
Oct 14, 2025
Sliding window attention (SWA) is powering frontier hybrid models for efficiency. Is there something better? Introducing Phalanx, a faster and better quality drop-in replacement for sliding window attention (SWA). Phalanx is a new family of hardware and numerics-aware windowed
12K
Radical Numerics
@RadicalNumerics
Oct 14, 2025
Replying to @RadicalNumerics
More on Phalanx and our research kernel library: Blog: radicalnumerics.ai/blog/phalanx Code: github.com/RadicalNumeric… Report: radicalnumerics.ai/assets/phalanx…
2.8K
Radical Numerics
@RadicalNumerics
Oct 14, 2025
Sliding window attention (SWA) is powering frontier hybrid models for efficiency. Is there something better? Introducing Phalanx, a faster and better quality drop-in replacement for sliding window attention (SWA). Phalanx is a new family of hardware and numerics-aware windowed
39K
Radical Numerics reposted
Michael Poli
@MichaelPoli6
Oct 11, 2025
We just released the largest open-source diffusion language model (RND1). RND1 is important to me on a personal level: it symbolizes our commitment to open-source exploration of radically different designs for AI at scale — training objectives, architectures, domains. There is
Radical Numerics
@RadicalNumerics
Oct 9, 2025
Introducing RND1, the most powerful base diffusion language model (DLM) to date. RND1 (Radical Numerics Diffusion) is an experimental DLM with 30B params (3B active) with a sparse MoE architecture. We are making it open source, releasing weights, training details, and code to
GIF
42K
Radical Numerics
@RadicalNumerics
Oct 9, 2025
Replying to @RadicalNumerics
Thank you @nebiusai, @PrimeIntellect, @LambdaAPI for the compute resources and support
6.6K
Radical Numerics
@RadicalNumerics
Oct 9, 2025
Replying to @RadicalNumerics
We’re also hiring aggressively. Reach out if you’re interested in building automated research environments and agents. (AI researchers and SWEs, pre/mid/post training, architecture design, kernels, lots of backend system design, and automation) Our team is behind the tech for
GIF
8.8K
Radical Numerics
@RadicalNumerics
Oct 9, 2025
Replying to @RadicalNumerics
More on RND1 models: Blog: radicalnumerics.ai/blog/rnd1 Code: github.com/RadicalNumeric… Report: radicalnumerics.ai/assets/rnd1_re… Weights: huggingface.co/radicalnumeric…
12K
Radical Numerics
@RadicalNumerics
Oct 9, 2025
Introducing RND1, the most powerful base diffusion language model (DLM) to date. RND1 (Radical Numerics Diffusion) is an experimental DLM with 30B params (3B active) with a sparse MoE architecture. We are making it open source, releasing weights, training details, and code to
GIF
848K