Ripple-65B Preview: a dense 65B agent model under Apache 2.0

Ricursive Intelligence, a chip-design lab, released Ripple-65B Preview on 4 November: a dense 65B text model for agentic coding with 1M context, Apache 2.0 and a 57 GB NVFP4 build.

  • New release
  • 2 min read
  • Published

Ricursive Intelligence, a US startup that uses AI to design chips, released Ripple-65B Preview on 4 November 2026 under the Apache 2.0 license. It is a text-only model with 64.8 billion parameters, aimed at agentic coding.

What is new

Ricursive calls Ripple a general-purpose variant of the proprietary model behind its chip-design platform. It was trained from scratch on 12 trillion tokens, mostly English text and code. It is a dense model, not a mixture of experts, so every parameter runs for every token, and Ricursive says that choice kept training simple enough for a team of one to five people. Of its 80 layers, 64 attend over a sliding window of 1,024 tokens and 16 use global attention, which keeps long-context memory in check. The card promises a context of 1,048,576 tokens.

In Ricursive’s own table it scores 80.5 on Terminal-Bench 2.1, ahead of the 80.1 it lists for Reflection Beam, 76.4 for Muse Spark Open, 63.8 for Inkling and 56.4 for Nemotron 3 Ultra. The same table has Beam ahead on SWE-bench Verified (81.6 against 76.5) and Muse Spark Open ahead on MCP Atlas (89.2 against 71.7). Ricursive notes that agentic scores shift with the test harness, and several rival numbers are its own in-house runs.

Running it

Four repositories are published: the BF16 instruction-tuned model (129.6 GB), a pretrain-only base model, an experimental NVFP4 build of 57.2 GB that quantizes only the MLP layers, and a speculative-decoding draft called DFlash2. The card says the NVFP4 build fits on an NVIDIA DGX Spark. By our estimate a 4-bit build needs about 45 GB, so a 48 GB GPU or a 64 GB machine with unified memory is the practical floor for local use. For servers, the repository ships an SGLang recipe for two B200 GPUs, where Ricursive reports 250 tokens per second at batch size 8 before quantization or custom kernels.

License

The weights use Apache 2.0: commercial use is allowed, with no revenue limit, no territory carve-out and no ban on using the outputs to train other models. The repository includes the LICENSE file, so the terms are easy to confirm.

Why it matters

Ripple-65B Preview shows that a dense model can report agentic-coding scores in the same range as much larger mixture-of-experts models, in a size that one multi-GPU node or a 128 GB desktop can serve. Reflection Beam, for example, has 501 billion parameters, and its smallest build is 294 GB. The cost is compute: all 64.8 billion parameters run for every token, against 23.4 billion active for Beam.

It is also a preview. Ricursive describes the post-training as relatively light, every benchmark number is self-reported, and the 1M-token context is its claim that we have not checked. For a smaller model with a similar self-reported score, see Ripple-65B vs Nanbeige4.3-33B, and to match a model to your hardware, read how to pick an open model for a coding agent.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Ripple-65B Preview Ricursive Intelligence 64.8B 1M Apache 2.0 (Permissive) 45 GB Workstation
Reflection Beam Reflection AI 501B 23.4B active 1M Apache 2.0 (Permissive) 323 GB Big-memory desktop
Muse Spark Open Meta 816B 42B active 1M Muse Model License (Conditional) 563 GB GPU server
Nanbeige4.3-33B Nanbeige LLM Lab 33.1B 512K Apache 2.0 (Permissive) 23 GB Gaming GPU