Ripple-65B Preview
Ricursive's Apache 2.0, 65B dense open-weight model with 1M context, claiming 80.5 on Terminal-Bench 2.1.
Overview
Ricursive Intelligence released Ripple-65B Preview on November 4, 2026: a 65B dense transformer pretrained on 12T tokens with up to 1M-token context, under Apache 2.0. Ricursive reports 80.5 on Terminal-Bench 2.1, 79.1 on SWE-bench Pro v2.0 Hard and 85.2 on GPQA Diamond, all self-reported, and trails on SciCode (39.4). The company states a $32M development cost, and the model is a general-purpose variant of the one behind its chip-design platform.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 65B dense (not MoE), 12T pretraining tokens, up to 1M-token context
- Apache 2.0 for all checkpoints
- Self-reported: Terminal-Bench 2.1 80.5, SWE-bench Verified 76.5, AIME 2025 98.3
- Hugging Face repos under ricursiveAI: instruct BF16, base BF16 and an experimental NVFP4 build
- Hosted at launch on OpenRouter and GMI Cloud at $0.20 input / $0.60 output per million tokens (16K context)
Use cases
- Self-hosted coding-agent backends where Apache 2.0 licensing matters
- Long-context agent work up to 1M tokens
- Local experiments with the NVFP4 build on DGX Spark-class hardware
Pricing
Free weights under Apache 2.0; hosted API listed at $0.20 input / $0.60 output per million tokens on OpenRouter and GMI Cloud (16K context).
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 4 Nov 2026
- Record updated
- 7 Nov 2026
- Record last checked
- 7 Nov 2026