Ricursive Intelligence reached target reinforcement learning performance for Ripple-65B Preview using 1.4 million rollouts

Overview Debates Related claims Topics Sources

Moonshot AI released its Kimi K3 model in July 2026, a 2.8-trillion-parameter mixture-of-experts system built to compete against artificial intelligence offerings from OpenAI and Anthropic

Nous Research released Arete-27B, an open-weight supervised fine-tune of Alibaba's Qwen3.8-27B designed to remove Chinese state framing from model responses

Nous Research trained Arete-27B using supervised fine-tuning on approximately 53 million tokens over two epochs without reinforcement learning or preference optimization

The training dataset for Arete-27B contains 34,878 prompts across seven value categories, with 29,236 accepted teacher answers forming the fine-tuning split

Based on the fine-tuning results of Arete-27B, Nous Research stated that the state alignment bias layer in Chinese language models is shallow and applied late in training

Thinking Machines pre-trained Inkling 2 Small on an NVIDIA GB300 NVL72 cluster containing 72 Blackwell Ultra GPUs per system in 68 hours, followed by nine days of reinforcement learning

Pipecat is an open-source framework that orchestrates voice and video streams while linking speech-to-text, large language models, and text-to-speech components to minimize latency in AI applications.

Brave Software cited in a blog post that Anthropic used its Search API to train the Claude 2 artificial intelligence model.

Deepgram expanded its Nova-3 speech recognition model to support 55 languages as part of its product strategy.

Baidu pre-trained ERNIE 5.5 Flash Base on 33.5 trillion tokens using the Muon optimizer with a 128,000-token training context window

ERNIE 5.5 Flash Base implements ECHO speculative decoding, which reuses the final backbone layer as a draft model adding 0.36 billion parameters excluded from headline counts

ERNIE 5.5 Flash Base features a hybrid attention mechanism across 43 layers, using Heavily Compressed Attention in 37 layers and Multi-head Latent Attention with sparse attention in 6 layers

ERNIE 5.5 Flash Base incorporates a tokenizer with 201,216 entries that groups numbers right-to-left in chunks of up to three digits

Ricursive Intelligence pretrained Ripple-65B Preview from scratch on 12 trillion text and code tokens at an 8,000 sequence length

Ricursive Intelligence reported requiring roughly 70 times fewer reinforcement learning rollouts for Ripple-65B Preview than the 100 million rollouts reported for Reflection Beam

Post-training for Ricursive Intelligence's Ripple-65B Preview incorporated synthetic data pipelines, long-context stages at 2 million sequence length, and reinforcement learning using truncated importance sampling

Ricursive Intelligence's Ripple-65B Preview features a dense Transformer architecture with 65 billion parameters across 80 layers and a model dimension of 5120

Pretraining of Ripple-65B Preview achieved roughly 40% model FLOPs utilization on NVIDIA hardware without custom kernels or quantization during training

The experimental NVFP4 checkpoint of Ricursive Intelligence's Ripple-65B Preview quantizes MLP layers while keeping attention in BF16 to target local hardware such as NVIDIA DGX Spark

Startup Ricursive Intelligence was founded by CEO Anna Goldie and CTO Azalia Mirhoseini, who previously helped create Google's AlphaChip chip design software