What Ricursive released
Ricursive Intelligence released Ripple-65B Preview last week. Together AI's model page lists 3 November 2026 as the release date. MarkTechPost covered the launch on 4 November. Ricursive is a Palo Alto lab that Anna Goldie and Azalia Mirhoseini founded in December 2025. Its main business is AI for chip design. Ricursive's announcement describes Ripple as a general-purpose release of the model behind that chip-design platform.
The weights are on Hugging Face under Apache 2.0. The release has four parts: a BF16 instruction-tuned model, an NVFP4 build of the instruct model, a BF16 base checkpoint and a 4.04B-parameter DFlash2 draft model for speculative decoding. In speculative decoding, a small draft model proposes several tokens at once and the large model checks them. Apache 2.0 allows commercial use, modification and redistribution. For a New Zealand team, that removes the licence questions that come with custom open-weight licences.
Ricursive's headline result is 80.52 on TerminalBench 2.1, an agentic coding benchmark. Reflection's 501B Beam scored 80.1 on the same chart. The model card also reports 76.5 on SWE-bench Verified and 85.2 on GPQA Diamond. Ricursive estimates its total R&D cost at about US$32 million. Every one of these numbers is self-reported. The model card says the scores were hard to replicate and calls its tables "a best-effort attempt at a fair comparison."
How the architecture works
Ripple is a dense Transformer. A dense model uses all of its parameters for every token. A mixture-of-experts (MoE) model sends each token through a small subset of its weights. Beam, for comparison, has 23B active parameters out of 501B. A dense 65B model needs far less memory to hold than a 500B MoE. Each token costs more compute and memory bandwidth than an MoE with 23B active parameters, though.
The model card describes a "thin-and-deep" layout with 80 layers and a model width of 5,120. Most layers use sliding-window attention, where each token sees only a recent window of earlier tokens. The vLLM recipe gives the pattern: four sliding-window layers with a 1,024-token window, then one full-attention layer that uses no positional encoding (NoPE). Only the full-attention layers keep keys and values for the whole context, so the KV cache grows much more slowly than in a model where every layer attends globally. The KV cache is the memory that holds earlier tokens' attention state. This design is what makes Ripple's 1,048,576-token context practical. The card also lists gated attention, value residuals and clipped feed-forward activations.
The training data is text only and mostly English and code, about 12 trillion tokens. Together lists a knowledge cutoff of 17 December 2025.
Running Ripple on your own hardware
At two bytes per parameter, the BF16 weights come to roughly 130 GB (our arithmetic). That's beyond almost any desktop. Quantisation stores weights at lower precision to shrink them. Third-party GGUF quantisations are much smaller: Toolhalla's weekly roundup puts a Q4_K_M file at 36.13 GiB. Even that won't fit on a single 24 GB GPU without heavy offloading to system RAM.
llama.cpp added the architecture in PR #31692, merged on 5 November. When Toolhalla checked on 6 November, the latest numbered release (b12144) didn't include the change, so early users had to build from master. The PR thread includes community test numbers. An AMD ROCm tester reported about 8.4 tokens per second at Q8_0 and 10.8 at Q6_K. A Vulkan user hit a buffer-size failure, and the author attributed it to Vulkan's storage buffer limit at contexts of 1M to 2M tokens.
Because Ripple is dense, each new token reads the full set of weights from memory. Memory bandwidth sets the speed ceiling more than raw compute does. That matters for the machine Ricursive names as a target, the NVIDIA DGX Spark. The DGX Spark has 128 GB of unified memory, which easily holds a 4-bit Ripple, but its memory bandwidth is 273 GB/s. A Spark will run Ripple comfortably for one developer. Don't expect it to serve a busy team.
DFlash2 speculative decoding should help with throughput, but llama.cpp support is still open as PR #31904. On an Apple M5 Max, the author measured 14.5 tokens per second without the drafter and 19.3 with four draft tokens. Reviewers asked why the gain was so small next to other DFlash2 models. One reviewer noted the PR doesn't handle the NoPE final layer.
On data-centre GPUs, the vLLM recipe serves the NVFP4 checkpoint on a single B200 or B300. NVFP4 is NVIDIA's 4-bit floating-point weight format. The recipe needs vLLM 0.33.0 or later, which currently means a nightly build with CUDA 13 wheels and an NVIDIA driver from the r580 series or newer. The recipe uses dedicated ripple parsers for reasoning and tool calls. Ripple decides on each turn whether to reason, so reasoning_content comes back empty when the model answers directly. The NVFP4 card keeps attention layers in BF16 and quantises only the MLP layers. The card's metadata lists 40B parameters, and the page doesn't explain the gap from 65B.
Hosted access and cost
If you want to test Ripple before buying hardware, Ricursive lists Together AI, OpenRouter, GMI Cloud and Fireworks as hosts. Together prices Ripple at US$0.20 per million input tokens, US$0.02 for cached input and US$0.60 per million output tokens. Together's page still showed a "we'll email you when the endpoint goes live" banner when we checked, though. Its spec table also lists a 128K context while the model card describes 1M. Check what context the host actually serves before you design around long inputs.
Limits
Ricursive calls this a preview with relatively light post-training focused on agentic coding. Even Ricursive's own charts show rival models ahead on maths, safety and automation tasks, so treat Ripple as a coding specialist. Nothing in the documentation covers te reo Māori. Given the English-heavy training data, test Māori-language output carefully before using it in a product. We also haven't found independent benchmark results yet.
What builders here could do with it
Some NZ teams can't send source code to offshore APIs because of client contracts, public-sector rules or plain caution. For them, a self-hosted coding agent on one 128 GB machine or a well-specced workstation is now a realistic option. Apache 2.0 also means teams can fine-tune Ripple on internal codebases and ship the result inside a commercial product.
A sensible first step is to run the Q4 GGUF or the hosted endpoint against tasks from your own repositories. Compare the pass rate with the model you use now, and measure tokens per second on the hardware you'd actually deploy. If you build a local Ripple setup worth showing, the society's projects page is open for write-ups.
Sources
- Ripple-65B Preview, a Frontier Open-Weight Model (Ricursive)
- ricursiveAI/Ripple-65B-Preview (Hugging Face)
- ricursiveAI/Ripple-65B-Preview-NVFP4 (Hugging Face)
- ricursiveAI/Ripple-65B-Preview-DFlash2 (Hugging Face)
- ricursiveAI/Ripple-65B-Preview vLLM serving recipe
- llama.cpp PR #31692: model : add Ripple (ricursiveai)
- llama.cpp PR #31904: Support DFLASH2 for Recursive Ripple-65B-Preview
- Ripple-65B Preview (Together AI)
- NVIDIA DGX Spark
- Ricursive AI Releases Ripple-65B Preview (MarkTechPost)
- Ricursive Intelligence Releases Ripple-65B Preview (OfficeChai)
- This Week in Local AI: 31 October to 6 November 2026 (Toolhalla)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.