Ricursive releases open Ripple-65B, claims top US score for coding agents

Ricursive Intelligence has released Ripple-65B Preview, an open-weight language model with 65 billion parameters that it trained from scratch. The company says the model scores 80.52 on Terminal-Bench 2.1, a test of coding agents working in a terminal. Ricursive calls that the state of the art among US open-weight models, ahead of rivals several times its size.

Ricursive describes itself as a full-stack AI lab that builds self-improving systems by co-optimising hardware, software and models. It trained the language model to power its own chip design platform. Ripple-65B Preview is a general-purpose variant of the proprietary model behind that platform. The release landed on 4 November 2026, within hours of Moonshot AI publishing the weights for Kimi K3.1 Pro, which Unicorn API covered separately.

What Ripple-65B Preview is

Ripple is a dense model. Every one of its parameters is used for every token it generates, unlike the Mixture-of-Experts (MoE) designs many labs use, which route each token through only a subset of the network. Ricursive says it chose a “modernized dense Transformer” to reduce infrastructure complexity in training and serving, and that the choice let it hold about 40% model FLOPs utilisation (MFU, the share of the hardware’s peak compute actually used) through pretraining and post-training. It did this with simple sharding, no custom kernels and no quantization.

The model was pretrained on 12 trillion tokens and supports a context window of up to 1M tokens. OrcaRouter, an AI model router that reviewed the release, adds that training ran 877,500 steps at an 8K sequence length on a text-only, mostly English-and-code corpus. A midtraining phase, a long-context stage and light reinforcement learning focused on agentic coding followed.

Spec	Ripple-65B Preview
Parameters	65B, dense (all active)
Layers / width	80 layers, d_model 5,120
Pretraining data	12 trillion tokens
Context window	Up to 1M tokens
Modality	Text in, text out
Licence	Apache 2.0
Weights	BF16 checkpoint plus NVFP4 and base variants
Terminal-Bench 2.1	80.52 (Ricursive’s evaluation)

The Terminal-Bench 2.1 claim and how it was framed

Terminal-Bench measures how well AI agents complete tasks inside a terminal, in reproducible containerised settings. The 2.0 version, described in a separate research paper, contains 89 curated tasks spanning software engineering, machine learning, security and data workflows. Ricursive’s result is on version 2.1. The sources do not describe how 2.1 differs from 2.0.

In Ricursive’s own evaluation, Ripple-65B Preview scores 80.52 on Terminal-Bench 2.1. The company lists these comparisons with larger US open-weight models:

• Reflection Beam 501B: 80.1
• Meta Muse Spark Open 816B: 76.4
• Thinking Machines Inkling 975B: 63.8
• NVIDIA Nemotron 3 Ultra 550B: 56.4

Two qualifiers matter. First, the claim is scoped to US open-weight models. Ricursive does not position Ripple against open models from other countries in its post. Second, the margin over Reflection Beam, which Ricursive calls the prior state of the art, is under half a point in its own numbers.

OrcaRouter says Ripple’s card also reports 79.1 on SWE-bench Pro v2.0 Hard, 44.3 on DeepSWE 1.1 and 76.5 on SWE-bench Verified. These are also the maker’s figures. None of the sources shows an independent leaderboard re-running Ripple.

How Ricursive says it trained the model for $32M

Ricursive says it developed and trained Ripple-65B Preview for $32M. It estimates that figure as one-tenth the cost of the previous open-weight leader on agentic coding, “to the best of our knowledge”. The company says the $32M is its full research and development cost from scratch. That includes pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, post-training recipe work and the final training runs. OrcaRouter describes the same figure as “the compute bill”. Readers should treat it as Ricursive’s all-in estimate, which no outside party has audited.

Post-training with fewer rollouts

Ricursive says it bought no post-training data. Instead it built its own synthetic data pipelines, quality filtering and coverage metrics. Its reinforcement learning (RL) recipe uses truncated importance sampling, score centering and a linear length penalty, with long-horizon synthetic environments. It added harder tasks as the model improved. The company says it reached 80.5 on Terminal-Bench 2.1 with 1.4M rollouts (attempts the model makes during RL training). It calls that roughly 70 times fewer than Reflection Beam used.

Architecture choices

The model uses a “thin-and-deep” shape: a narrow 5,120-wide residual stream and 80 layers. To improve reasoning, Ricursive overprovisions attention heads. To offset the cost, it raises the ratio of sliding-window attention (SWA) layers to global layers and shrinks the SWA window. SWA layers only look at nearby tokens. The design also adds FFN activation clipping, gated attention, global NoPE (no positional encoding in global layers) and value residuals. OrcaRouter characterises the result as a move away from the Olmo 3 architecture toward one resembling Gemma 3N.

Ricursive says it tuned its pretraining recipe through “model laddering”, a series of small-scale experiments, starting from Olmo 3’s ladder. Some changes were kept. A switch to pre+post normalisation was neutral at small scale but helped from 4B parameters upward. Others were rejected: a 2-stream mHC variant looked good small but did not improve at 4B. The company reports that the combined changes, plus a custom dataset with heavily upsampled code, improved compute efficiency on code by up to 16 times at 4B scale. That means matching the Olmo 3 ladder’s code performance with 16 times less compute, as measured by dolma_stack cross-entropy. Ricursive argues the gain should grow with scale, because the slopes of the two ladders differ.

Where Ripple-65B is available

Ricursive says Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks going live later on launch day. OrcaRouter reports the licence as Apache 2.0, a permissive licence that allows commercial use. It says the weights ship as a single-repository BF16 checkpoint plus NVFP4 and base variants.

Ricursive also pitches local use. It says the model’s relatively small parameter count lets it run efficiently on devices such as NVIDIA DGX Spark. That machine has high compute but low memory bandwidth compared with datacentre accelerators. The company adds that the dense design gives low-latency token generation.

Why it can’t be ranked against Kimi K3.1 Pro

The same-day release of Kimi K3.1 Pro invited side-by-side comparisons. OrcaRouter argues these comparisons are not valid. Kimi’s launch reports 65.7 on Terminal-Bench 4.0, which OrcaRouter describes as “a later, harder revision of the harness than 2.1.” Setting that against Ripple’s 80.5 on 2.1 compares two different exams. OrcaRouter counts six coding benchmarks across the two launches, with not one of them shared.

OrcaRouter adds that scores vary even on a single benchmark. Kimi K3.1 Pro’s Terminal-Bench 4.0 result appears as 65.7 from Moonshot, 63.6 from Vals AI and 67 from Artificial Analysis. That is a 3.4-point spread on the same model and version, which OrcaRouter attributes to tooling. That spread is wider than Ripple’s stated lead over Reflection Beam on 2.1, although those numbers come from different benchmarks and harnesses.

The two models are also built differently. OrcaRouter puts Kimi K3.1 Pro at about 2.8T total parameters with roughly 104B active per token, citing third-party trackers. It describes Kimi’s weights as about 1.56 TB, under a renamed Kimi K3 licence with a $20M/12-month MaaS threshold, and says Kimi accepts text, image and video input. Ripple is text-only and under Apache 2.0. “One is small enough to sit on a desk; the other needs a datacentre,” OrcaRouter writes.

What it means for builders, and the open questions

For teams that want an open coding-agent model they can host themselves, Ripple offers several things together. Its terminal-agent score is, by Ricursive’s count, the best among US open-weight models. It uses a single dense architecture that Ricursive says needed no custom kernels to train. And it comes under a permissive licence. A dense 65B model is also simpler to serve than a trillion-parameter MoE, and Ricursive explicitly targets desk-side hardware.

The limits are also clear from the sources:

• Self-reported scores. The Terminal-Bench 2.1 and SWE-bench figures come from Ricursive. OrcaRouter’s analysis shows that harness differences alone can move a score by several points.
• Preview status. The model is labelled a preview. It is a general-purpose variant, while an internal variant powers Ricursive’s chip design platform.
• Narrow inputs. It is text in, text out, trained mostly on English and code, so multilingual and multimodal work are outside its stated scope.
• Cost claim. The $32M figure and the “one-tenth” comparison are Ricursive’s own estimates.
• Scoped comparison. The state-of-the-art claim covers US open-weight models only.

What to watch next

The next step is independent evaluation. Third-party runs of Terminal-Bench 2.1 and the SWE-bench variants would show whether Ripple’s narrow lead over Reflection Beam holds outside Ricursive’s own setup. A shared benchmark version with Kimi K3.1 Pro would allow a real comparison. Also worth watching: whether the “Preview” label gives way to a full release, how the Fireworks listing and other hosts perform in practice, and whether Ricursive publishes more about its chip design platform. The company says it built Ripple “to begin closing the loop between models and the hardware that fuels them”.