Tasks and inputs
Ripple-65B Preview is listed for Text generation, Coding & Code Generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Ripple-65B Preview is Ricursive Intelligence's first open model: a dense 65B coding-focused transformer trained from scratch, with a 1M-token context.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Dense models at this size have become rare. Most new open releases above 50B parameters route through experts; Ripple-65B Preview, released on 4 November 2026 by Ricursive Intelligence, does not. Every one of its 65B parameters fires on every token, it was pretrained from scratch rather than distilled from someone else's model, and it reads up to 1,048,576 tokens.
It is Ricursive's first open model, licensed Apache-2.0.
The shape is unusual: a narrow model dimension of 5,120 stretched over 80 layers. Attention mixes a high ratio of sliding-window to global layers, gated attention, global layers without positional encoding (NoPE), value residuals and FFN activation clipping. Ricursive moved to this layout from an earlier Llama-style design for training stability.
Pretraining used 12T tokens, mostly English and code, at 8K sequence length. Long-context stages then stretched sequences to 2M tokens, followed by SFT and a short RL run. Post-training was deliberately light and aimed at agentic coding.
| Benchmark | Ripple-65B Preview |
|---|---|
| Terminal-Bench 2.1 | 80.5 |
| SWE-bench Verified | 76.5 |
| SWE-bench Pro v1.0 | 61.42 |
| GPQA Diamond | 85.2 |
| MMLU-Pro | 84.2 |
| IFBench | 79.2 |
Treat these with care, for three reasons Ricursive itself raises or that observers have flagged. Agentic scores are hard to replicate, and the team calls its comparison best-effort. Several competitor figures are Ricursive's own runs rather than official numbers. And Ricursive's own published tables and launch graphic disagree on some results, including a SWE-bench Pro v2 Hard figure. Terminal-Bench also came out lower under the official harness with summarisation at 128K.
Even so, the base checkpoint is credible on classic tests: MMLU 83.7, GSM8K 88.5, HumanEval 70.9.
Dense 65B in BF16 is heavy. Ricursive quotes about 250 tokens per second on two B200s at batch 8, before any quantisation or custom kernels, and says the design targets machines like the DGX Spark. An NVFP4 checkpoint is published for that route. A pretrain-only base and a DFlash2 variant are also out.
Hosting is unsettled. Secondary reports list several inference providers, but at the time of writing none showed on the model's own page. Expect to self-host, and confirm any provider listing yourself.
You want a large open coding model without MoE serving complexity, or you need million-token context in a model you control. Compare it against DeepSeek V4 Flash and Nemotron 3 Ultra on your own repositories.
A useful test: take five long-lived repositories, feed entire directories into context, and ask for cross-file refactors with tests. Measure pass rate and peak memory together. Tool-calling and reasoning-mode behaviour are not documented in detail, so validate your agent harness's tool format before a full run. More options sit under coding models.
Release dates vary by a day across trackers (4 or 5 November). The repository is the same.
Compare the model on the work you need to complete.
Ripple-65B Preview is listed for Text generation, Coding & Code Generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 1.05M context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Microsoft
MAI Code 2 is Microsoft AI's coding model for multi-file work, debugging and frontend builds, live in GitHub Copilot at $0.50/$2.50 per 1M (promo).
Microsoft
MAI Code 2 Flash is Microsoft AI's low-cost coding model for edits, fixes and explanations, on every GitHub Copilot plan including Free and Student.
Microsoft
MAI Cyber 2 Flash is Microsoft AI's security model for vulnerability discovery and threat investigation, with a 1M-token context, shipping in MDASH.
Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agentic execution, with a 1M context, xhigh thinking and a free tier.
OpenAI
GPT-6 Cyber is OpenAI's restricted security model for authorised vulnerability investigation, reproduction and fix verification, for Daybreak Red customers.
particle-ai
Particle-1 is Particle's first software-engineering model, an RL-tuned DeepSeek V4.1 Flash that aims to solve coding tasks with fewer tokens.