Frontier code agents need to run long tasks in a terminal without losing track, and pretraining them is usually billed in hundreds of millions. That is the gap Ricursive says it built Ripple-65B Preview to close, a 65-billion-parameter model pretrained on 12 trillion tokens with support for up to 1M context length.
Ripple-65B Preview is a dense Transformer, an architecture that activates all parameters for every token, and Ricursive says the general-purpose variant of the proprietary model that powers its AI chip design platform sets a new U.S. open-weight state of the art at 80.52 on TerminalBench 2.1, a benchmark that scores agents on executing real coding tasks inside a Linux terminal.
That score beats significantly larger U.S. open models on the same benchmark.
Ricursive puts Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4, according to the company’s release. The bet is that a smaller dense model can win on latency and local deployability where sparse models need more memory. Ripple-65B Preview is built to fit devices like NVIDIA DGX Spark, which Ricursive notes has high FLOPs with low memory bandwidth versus datacenter accelerators.
How Ricursive says it got there starts before posttraining. Ripple-65B Preview uses model laddering, a practice of running systematic small-scale experiments to pick architecture and data mixes before a single large run, starting from the Olmo 3 ladder from the Allen Institute for AI. Ladder v1.0 added StepFun batch size and learning rate scaling laws, v1.1 scaled depth and width, v1.2 moved from post-norm to pre+post norm, v1.3 added sigmoid gated attention, a mechanism that learns a gate to control how much attention output flows forward, v1.5 switched to global NoPE, a technique that removes explicit positional embeddings and relies on other signals for position, plus a final learning-rate cooldown to zero, and v1.6 added value residuals, a shortcut that carries value vectors forward to aid reasoning. Ricursive rejected a 2-stream mHC variant that helped at small scale but not at 4B, and said the final ladder plus a custom dataset with heavily upsampled code improved compute efficiency on code, as measured by dolma_stack CE, by up to 16x at 4B scale versus the Olmo 3 baseline.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
The architecture stays dense by design. Ricursive chose not to use Mixture-of-Experts, a technique that routes tokens to a subset of expert sub-networks to increase parameter count without increasing per-token compute, to keep sharding simple and avoid custom kernels or quantization while holding about 40% model FLOPs utilization. Ripple-65B Preview is thin and deep with a narrow d_model of 5120 and 80 layers, overprovisions attention heads for reasoning and offsets that cost by raising the ratio of sliding window attention, a method that restricts attention to a recent window, to global attention layers and shrinking the window size. The build also adds FFN activation clipping, a method that bounds feed-forward activations to stabilize training, alongside gated attention and the NoPE and value residual changes noted above.
Posttraining is where Ricursive claims the steepest gain. The team did not buy posttraining data and instead built synthetic data pipelines with quality filtering and coverage metrics. Supervised fine-tuning, the stage that teaches instruction following on curated examples, provides the starting checkpoint, then a mixed reinforcement learning recipe that uses truncated importance sampling, a correction that caps how much off-policy data influences updates, plus score centering and a linear length penalty tuned long-horizon synthetic environments. Ricursive says that stack climbed to 80.5 on TerminalBench 2.1 in only 1.4M rollouts, roughly 70x fewer than Reflection Beam as the prior state of the art, and that pretraining plus posttraining together cost $32M end-to-end including ladder development, data mixing, evals, debugging, synthetic data and the final hero runs, which the company estimates is one-tenth the prior open-weight state-of-the-art cost on agentic coding.
Ripple-65B Preview is available now on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks slated to go live later, and Ricursive says on X that Ripple-65B Preview is a dense 65B model pretrained from scratch for agentic coding. For builders, the artifact to test is whether that ladder and RL recipe reproduce the TerminalBench lead on your own terminal workloads, and whether the dense 65B envelope actually delivers the promised low-latency and local-device efficiency without MoE tricks.
