
Founding Technical Staff @RicursiveAI 's LLM team AlphaChip | AI for Chip Design | ex-@GoogleDeepMind, ex-Google Brain
Palo Alto, CA
Joined November 2025
- Tweets95
- Following1.1K
- Followers553
- Likes7.7K
Ebrahim Songhori retweeted
Really enjoyed particpating in the @RicursiveAI hackathon with @seilk_ and @hangoo_kang .
We showed how Ripple was capable of building and operating its own code IDE from scratch, achieving < 80 ms autcomplete and 300 tok/s code generation.
The first @RicursiveAI hackathon, centered on Ripple-65B Preview, was so much fun! Participants built impressive computer use demos, a lightning-fast IDE built and operated by Ripple, kernel designs that outperformed Luna, and more.
We never directly optimized the model for some of these capabilities, like some cool visual demos.
Ebrahim Songhori retweeted
Last Friday we hosted the Ricursive Intelligence After Dark Hackathon, our first hackathon centered on our open-weight model release, Ripple-65B Preview. We brought together an incredible group of researchers, engineers, and builders to push the model in new directions. We saw some amazing projects spanning everything from CUA and Weight-Space Programming to model watermarking.
Huge thanks to everyone who came out, built something, and made the night memorable! This is just the beginning!
Ebrahim Songhori retweeted
Last week, we released Ripple-65B Preview. A 65B dense LLM that set a new U.S. open-weight SOTA on agentic coding.
Join our very own @DZhang50, Founding Member of Technical Staff, and @togethercompute's @zainhas to learn how we did it.
📆 Tuesday, Nov 17 · 11am PT / 2pm ET
👉 Register at luma.com/together-1kx8
How do you build a small coding model from scratch?
@RicursiveAI 's Dan Zhang breaks down the architecture and decisions behind Ripple-65B Preview with @zainhas: dense over MoE, the ablations that held up (and didn’t), and what happened inside the RL runs.
Tuesday, Nov 17 . 11 A.M. PT / 2 P.M. ET
Ebrahim Songhori retweeted
Last week we released Ripple-65B Preview. This Friday, come build with it.
Ricursive After Dark Hackathon: Fri Nov 13, 3PM–10PM, at our Palo Alto office. Compute, dinner, and hardware prizes.
Applications close Thursday, November 12 at noon PT. Accepted participants will be notified by email Thursday night with the address and full details.
luma.com/qnsar3a2?utm_source…
Major speed breakthrough: Ripple-65B-Preview running at up to 124 tok/s with concurrency on a single DGX Spark 👇
One stream already ran at 24 tok/s on greedy code. But eight streams were stuck at 49. I had a hunch that there was substantial potential for improvement. So I got my agents to work.
The drafter guesses several tokens ahead, and the model checks them all in one pass. When 8 streams shared a round, the engine got the cost of each stream's drafts wrong, so each stream drafted about 1 token a round. Now each drafts about 8, and on greedy code the model keeps about 6 of them.
At 8 streams, each stream now runs at 15.5 tok/s, up from 6.1.
At 8 streams, total tok/s, before → after the fix:
• Code, greedy: 49 → 124 (2.5x)
• Code, sampled: 35 → 60
• Prose, greedy: 41 → 62
• Prose, sampled: 34 → 52
The speed-up is only as good as the drafter. It guesses best on greedy code, where the next tokens are the most predictable, so that is where the gain is largest.
Every drafted reply equals the plain reply, token for token.
It points to what's possible on a single Spark: a large dense model where only the MLP weights are 4-bit (NVFP4). Attention, embeddings and the output head all run in BF16.
Setup: my CUDA code for the DFlash2 drafter for TensorFold, on GB10, with @RicursiveAI's drafter weights. 256-token replies, median of 3 runs, both builds in one session. Every stream sends the same prompt, so the 8 greedy streams write the same reply, but drafts are not shared between streams. Sampled streams (T=1.0) have their own seeds.
cc @ashxhart @andersjw_ @MiaAI_lab @bertholomusai @CerebralCoding_ @NVIDIAAI @NVIDIARTXSpark may be of interest to you 💜
📢 Ripple 65B is live on @FireworksAI_HQ and @togethercompute
Now live on Fireworks and Together as well! All inference endpoints available below:
Fireworks: fireworks.ai/models/firework…
Together: together.ai/models/ripple-65…
GMI Cloud: console.gmicloud.ai/model/ll…
OpenRouter: openrouter.ai/ricursiveai/ri…
1
11
650
Very good and detailed review of our model from @Sam_Witteveen
Ripple 65B: A Chip Design Lab Just Beat Every US Open Model youtu.be/2dD4HXg-YJ4?is=BTHp…
Ebrahim Songhori retweeted
Ripple-65B from @RicursiveAI is a good model. We tested it ourselves on one RTX PRO 6000.
After warmup, in our agent: steady 50-70 tok/s on prose ("write an essay and a poem about regret") and 110-120 on code ("write a simple python script 1000 lines").
#LocalLLM #OpenSourceAI
Ebrahim Songhori retweeted
I made a 3D visualization of Ripple-65B Preview's architecture! Check it out here (with an easter egg): shayantalaei.com/ripple-65b/
Ripple is a dense model with 80 layers and a width of 5,120, so its aspect ratio is 64. That makes it very deep for its width. In a model this deep, each layer keeps adding its output to the residual stream, and by the deeper layers the original token information gets diluted.
Most recent architectures handle this by changing the residual stream itself, like DeepSeek's manifold-constrained hyper-connections (mHC) or Kimi's Attention Residuals. We went with value residuals instead. Every layer mixes the original token embedding back into its attention values, and a learned weight in each layer decides how much.
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
First Ripple-65B GGUF 🔥
Nothing could run Ripple locally, so we ported it to llama.cpp. Checked against @RicursiveAI's reference code: the port matches its token choice 100% of the time.
Q8_0 68.8 GB · Q4_K_M 38.8 GB on disk
24/32 GB quants coming
huggingface.co/AtomicChat/Ri…
Ebrahim Songhori retweeted
Update: the OpenRouter link is here: openrouter.ai/ricursiveai/ri…
Ebrahim Songhori retweeted
first model! we built a SOTA U.S. open weight model at 1/10th the cost
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
Access @RicursiveAI 's model here:
HuggingFace: huggingface.co/ricursiveAI/R…
Cloud: console.gmicloud.ai/model/ll…
OpenRouter: openrouter.ai/gmicloud/rippl…
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
Congrats to @RicursiveAI on Ripple-65B Preview, their first open model 🎉
It was trained on NVIDIA B300 GPUs & sharpened with open NVIDIA Nemotron post-training datasets.
Download the NVFP4 checkpoint from Hugging Face and run locally on DGX Spark 👉 huggingface.co/ricursiveAI/R…
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
Exciting - OSS LLM based in the US!!
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
GMI Cloud is proud to be a Day 0 launch partner for Ripple-65B Preview from @RicursiveAI 🚀
It's a 65B dense model pretrained from scratch, and it sets a new state of the art for U.S. open-weight models on agentic coding tasks, ahead of much larger models.
Try it now 👇
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
happy to have posttrained what was known as Rizz-65B up until an hour ago
shoutout to the team:
@talaei_shayan @DZhang50 @ethanboneh @esonghori2
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple
Ebrahim Songhori retweeted
Ok, *this is impressive*: Ricursive Intelligence just released Ripple-65B Preview, an open-weight AI model that it says outperforms much larger US models on a key agentic coding benchmark. And even runs locally on my DGX Spark.
Ripple scored 80.5% on Terminal-Bench 2.1, narrowly ahead of Reflection’s Beam-501B at 80.1%, according to the companies’ reported results!
The size comparison comes with a detail: Ripple uses all 65B parameters for each task, while Beam activates 23B of its 501B parameters.
Ricursive says Ripple was trained on 12 trillion tokens with a $32 million compute budget - from scratch. The model supports up to 1 million tokens of context and is available under Apache 2.0.
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: ricursive.com/ripple



























