Stephen Ebert retweeted
ZHA is growing up!
We’ve scaled @ZyphraAI’s subquadratic attention method - ZHA - to leading open source multi-trillion parameter models across Nemotron, Qwen, Mimo, and GPT-OSS!
ZHA preserves reasoning and agentic performance while delivering 9-60x long-context speedups at 10M tokens.
GLM-5.3-Flash (FP4) is on Zyphra Inference at 25% below our FP8 price, served with the Zyphra Optimization Engine (ZOE) on @AMD Instinct MI300X.
With ZOE, p90 generation is 9.5x faster per user and p90 TTFT is 48x lower than stock SGLang.
Try it at: cloud.zyphra.com
@ZyphraAI is on the @ArtificialAnlys leaderboard for GLM-5.3-Flash throughput. Optimized on @amd MI300X with Zyphra’s Optimization Engine (ZOE).
Try it on Zyphra Inference: cloud.zyphra.com
$/1M tokens:
- Input: $0.15
- Cached: $0.03
- Output: $0.25
Congrats to @Gcabrielclark and team! 🙌
Meet ZONOS3-Mini: open-source text-to-speech with voice cloning and word-level control. All on-device and super fast.
Every voice in this video is from a TTS model that can run real-time on a phone.
Comment “ZONOS” within 2 days for 20 hours of free audio.
Sound on 🔊
5
120
Stephen Ebert retweeted
We were finding some oddities while trying to improve FP8 attention. How come improving attn numerics worsens speed? It's actually pretty simple: bad numerics cause underflows to zero, zero values don't flip as much, flipping fewer bits requires less current, less current allows higher clock speed. @cHHillee thought strikes again (thonking.ai/p/strangely-matr…). Keeping per-cycle clock speed high, absolute number of clock cycles low, and numerics accurate was hard to juggle!
Introducing Zyphra FlashAttention (ZYFA): faster, more accurate 8-bit floating-point (FP8) attention on AMD MI355X.
Our prefill kernels deliver 1.16–1.76x the throughput of AMD’s inference kernel library, AITER, on real model inputs at 16K tokens under sustained load. 🧵
GLM-5.3-Flash from @zai_org is on Zyphra Inference, served with the Zyphra Optimization Engine (ZOE) on @AMD Instinct MI300X.
With ZOE, p90 generation is 3.1x faster per user and p90 time to first token is 10.7x lower than stock SGLang.
Try it at: cloud.zyphra.com
Today we’re introducing Zyphra Hierarchical Attention (ZHA), a path to true subquadratic attention.
ZHA searches a learned tree of compressed summaries, with projected speedups of 35× over dense attention and 5× over MSA at 10M tokens, increasing to ~400× and ~55× at 1B. 🧵
Zyphra presents the second wave of inference optimizations for @AMD Instinct MI355X GPUs, powered by Zyphra Optimization Engine (ZOE).
Our latest engineering blog details the kernel work behind faster long-context serving for frontier open models. 🧵
Stephen Ebert retweeted
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Stephen Ebert retweeted
My new video on ai and design is now live on @Microsoft’s official website!
I believe designing intelligence means shaping humanity.
In the video, I talk about:
- Building @copilot into a thinking partner
- Designing beyond interfaces: models, memory, evaluation...
And a question I keep coming back to: as ai reshapes our world, how can we live fully, creatively, and meaningfully?
Stephen Ebert retweeted
Gemini 4 Argon from @GoogleDeepMind is live on OpenRouter, via Google AI Studio and @googlecloud.
It tops AutomationBench (51.3%) and LVBench (91.7%) and ties for first on CWE-bench v1 (68%).
$2/M input, $10/M output (introductory pricing).
Try it now: openrouter.ai/google/gemini-…
In Mixture of Expert (MoE) models the cost of moving tokens to their experts can dominate total runtime.
At Zyphra research, we use patterns in how tokens are routed to experts to make that communication faster by up to 2.63x on @AMD MI300X GPUs, with the model unchanged.
Stephen Ebert retweeted
Many great results from OpenAI. Congratulations!
This one in particular is very close to my heart. A decade or so ago, with collaborators, I have published a physics-oriented proof of it.
github.com/openai/math/blob/…
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
How do language models develop a sense of time without explicitly encoding position in their attention operations?
Zyphra Research explains how local memory layers that read nearby words help global attention layers keep track of word order. 🧵
Stephen Ebert retweeted
Excited to share my new result on unbounded Holevo additivity gaps, a long-sought open question since Hastings’ seminal discovery on addivitiy violation.
We identify a sequence of channels where Holevo $\chi$ diminishes while classical capacity diverges. Surprisingly, the construction was hiding in plain sight: it simply takes a fresh look at earlier work by Benoît Collins et al. bridging random channel violations and free probability theory (, which is also a subject of my favorites, and more works coming soon, stay tuned!).
arxiv.org/abs/2609.18222

























