Ricursive Intelligence has released Ripple-65B Preview, an open-weight model designed specifically for agentic coding. The system scores 80.52 on Terminal-Bench 2.1, a figure the company claims is the highest among US-based open-weight models.
Every Mixture-of-Experts rival that beats this score on the same test carries between 501B and 975B total parameters. Ripple-65B Preview is a dense architecture, pre-trained from scratch to handle complex coding tasks.
Specification summary
- The model is 65B dense, pre-trained on 12T tokens, with a context window of up to 1M.
- It runs on BF16 using datacenter GPUs, achieving 250 tokens per second on two B200 chips. An experimental NVFP4 build targets NVIDIA DGX Spark.
- Performance matches frontier-level agentic coding for its size, though it trails larger MoE models on knowledge and science coding.
- Best result: 80.5 on Terminal-Bench 2.1.
- Worst result: 39.4 on SciCode, the lowest score among the five models in Ricursive’s own comparison table.
The release package
The download includes three checkpoints: an instruction-tuned BF16 model, an NVFP4 version, and a BF16 base model. All ship under the Apache 2.0 license. Pretraining used a text-only corpus, mostly English and code, at an 8K sequence length.
Ricursive states the full research and development cost was $32M. That figure covers ladder experiments, data mixing, evaluations, synthetic data and the final runs. The team estimates this is one-tenth the cost of the prior open-weight agentic coding leader.
Architecture details
Ricursive chose a dense Transformer over Mixture-of-Experts to reduce infrastructure complexity. Training held 40% MFU with simple sharding, no custom kernels and no quantization.
The model is thin and deep: a d_model of 5120 across 80 layers. It moves from Olmo 3’s Llama-like layout to a Gemma 3N-like design for stability. Key choices include:
- Overprovisioned attention heads for reasoning, offset by more sliding-window layers and smaller windows.
- Gated attention, global NoPE and value residuals.
- FFN activation clipping, a technique also used in gpt-oss.
Ricursive built the recipe through laddering experiments starting from Olmo 3’s ladder. It rejected changes that failed to hold at 4B scale, such as 2-stream mHC. The technical post reports a 16x code compute-efficiency gain at 4B scale.
Post-training process
Post-training ran midtraining, long context, SFT, long-context SFT and mixed RL. Long-context stages trained at 2M sequence length. Ricursive bought no post-training data and built its own synthetic pipelines.
The RL recipe combines truncated importance sampling, score centering and a linear length penalty. Ripple hit 80.5 on Terminal-Bench 2.1 with 1.4M rollouts. Ricursive puts that at roughly 70x fewer than Beam, which reported over 100M rollouts.
Benchmark performance
Ricursive ran missing evaluations in-house using OpenCode, and Terminus 2 for SWE-bench Pro v2.0 Hard and DeepSWE 1.1. Scores below are Ricursive’s presented data.
- Terminal-Bench 2.1: 80.5, versus Beam 80.1, Muse Spark Open 76.4, Inkling 63.8 and NVIDIA Nemotron 3 Ultra 56.4.
- SWE-bench Pro v2.0 Hard: 79.1, the top score in Ricursive’s table.
- SWE-bench Verified: 76.5, below Beam and Inkling.
- Math: 98.3 on AIME 2025 and 90.5 on HMMT Feb 2026.
- Weak spots: SciCode 39.4, DeepSWE 1.1 44.3, MCP Atlas 71.67, GPQA Diamond 85.2.
Ricursive also reran Terminal-Bench 2.1 with the official Terminus 2 harness at 128K. Most rivals dropped there; Beam fell to 70.8. Muse Spark Open rose to 78.7, still below Ripple’s 80.5.
Comparison table
| Feature | Ripple-65B Preview | Reflection Beam | Meta Muse Spark Open | Thinking Machines Inkling | NVIDIA Nemotron 3 Ultra |
|---|---|---|---|---|---|
| Architecture | Dense | MoE | MoE | MoE | Hybrid Mamba-2 LatentMoE |
| Total params | 65B | 501B | 816B | 975B | 550B |
| Active params | 65B | 23B | 42B | 41B | 55B |
| Pretraining tokens | 12T | 23.8T | Not disclosed | 45T (per Ricursive table) | ~20T |
| Context | Up to 1M | 1M (midtraining) | 1M | 1M | Up to 1M |
| License | Apache 2.0 | Apache 2.0 | Muse Model License | Apache 2.0 | OpenMDW-1.1 |
| Input modality | Text | Text | Text, image, video, audio | Text, image, audio | Text |
| Self-hosting floor | DGX Spark class (NVFP4) | Not disclosed | 700GB VRAM (NVFP4) | 600GB VRAM (NVFP4) | 8x B200 or 16x H100 |
| Terminal-Bench 2.1 | 80.5 | 80.1 | 76.4 | 63.8 | 56.4 |
| SWE-bench Verified | 76.5 | 80.9 | Not disclosed by Meta | 77.6 | 70.7 |
| GPQA Diamond | 85.2 | Not disclosed in blog | 89.0 | 87.2 | 87.0 |
Ripple scores from the Ricursive model card. Competitor scores from each vendor’s own source. All self-reported.
Deployment options
Weights are on Hugging Face. Ricursive lists Together AI, GMI Cloud and OpenRouter at launch, with Fireworks later. Together AI’s page currently marks the model as coming soon.
On datacenter GPUs, the model reaches 250 tok/s on 2x B200 at batch 8, before quantization. The experimental NVFP4 checkpoint quantizes MLP layers and keeps attention in BF16. Ricursive positions it for local devices like NVIDIA DGX Spark. A DFlash2 variant is also listed.
What it means
Developers can now run a dense 65B model on local hardware that previously required massive MoE clusters. The 80.5 score on Terminal-Bench 2.1 suggests strong performance for agentic coding tasks. However, the 39.4 result on SciCode indicates significant weaknesses in science-specific coding compared to larger rivals.
