The Short Answer
Ripple-65B Preview is a dense 65B open-weight coding model under Apache 2.0 that reports 80.5 on Terminal-Bench 2.1. The scores are self-reported and the lead is narrow, but the license is real: you can run it inside your own network. On ibl.ai you own all the code and the data and run agents model-agnostic, so a model like this is a setting.
Ricursive Intelligence released the model on November 4, 2026, with weights and a model card on Hugging Face. VentureBeat's review is the most careful independent read so far.
What is Ripple-65B and who built it?
Ripple-65B Preview is a 65-billion-parameter dense language model built for agentic coding by Ricursive Intelligence. The model card says it was trained from scratch on 12 trillion tokens and supports a context of about one million tokens.
Ricursive's CEO Anna Goldie and CTO Azalia Mirhoseini previously helped create Google's AlphaChip, per VentureBeat. The company says the full build cost about $32 million in compute.
It ships as BF16 and compressed NVFP4 instruction-tuned checkpoints plus a BF16 base model, all under the Apache 2.0 license.
Does Ripple-65B really beat a 501B model on Terminal-Bench 2.1?
On Ricursive's own numbers, yes, by 0.4 points: Ripple-65B reports 80.5 on Terminal-Bench 2.1 against 80.1 for Reflection's Beam-501B. Meta's Muse Spark Open is listed at 76.4 and NVIDIA's Nemotron 3 Ultra at 56.4.
Two caveats travel with that. The figures are self-reported, and the model card mixes vendors' published results with Ricursive's in-house runs. It also says Terminal-Bench 2.1 scores were lower under the official Terminus 2 harness for every model except Muse Spark Open.
Under Terminus 2, the card lists Beam at 70.8 and Muse Spark Open at 78.7, with Ripple still at 80.5. Ripple's lead holds on that harness too, but the comparisons are best-effort, not a controlled head-to-head.
Is "eight times smaller" the right way to read Ripple-65B against Beam?
Not quite. Beam-501B is a mixture-of-experts model with about 23B parameters active per token, per the model card, while Ripple is dense, so all 65B are active. On compute per token, Ripple does more work than Beam, not less.
Where size does matter is memory. A 65B dense model is small enough that Ricursive cites NVIDIA's DGX Spark as a local target, and the card lists 250 tokens per second on two B200 GPUs at batch size 8.
BF16 weights for 65B parameters come to roughly 130 GB, which suggests the NVFP4 checkpoint is the practical fit for a single desktop box.
Where does Ripple-65B fall short?
Ripple-65B leads only three of the 22 rows in its main results table, by our count of the model card's table: the two Terminal-Bench rows and SWE-bench Pro v2.0 Hard, where it reports 79.1 against Beam's 76.5.
It trails elsewhere. Muse Spark Open reports 60.0 on DeepSWE 1.1 to Ripple's 44.3, and 94.2 to 83.8 on the FORTRESS adversarial safety set. Ricursive calls this a Preview with light post-training.
So the accurate summary is a strong agentic coding specialist, not a general frontier model, and a claim that needs independent evaluation before anyone builds a procurement case on it.
Why does an Apache 2.0 coding model matter to an enterprise?
Apache 2.0 matters because it removes the legal question from self-hosting. The license allows commercial use, modification and redistribution, so a team can run the model on its own hardware with no usage fee or field-of-use limit to negotiate.
For coding agents that is the point. An agent working on a private repository sends that code to wherever the model runs. A model you host keeps it inside your network.
Price pressure is the second effect. OpenRouter lists Ripple at $0.20 per million input tokens and $0.60 output through GMICloud, per VentureBeat. Our piece on open weights as a price floor covers how that pressure reaches closed-model pricing.
Should an enterprise switch its coding agents to Ripple-65B now?
Not on a self-reported benchmark alone. A narrow lead on one evaluation is the kind that changes hands within weeks, and the model card itself says agentic scores are hard to reproduce.
The better move is to be able to test it cheaply. Run it on your own tasks, next to the model you use today, and compare cost and output. That only takes an afternoon if your agents name their model as a setting.
Revolut showed what that looks like at scale: it moved its AI coding to self-hosted open models and reported costs at roughly a tenth of frontier-lab prices.
How does ibl.ai make a model like Ripple-65B usable?
On ibl.ai you own all the code and the data, the platform runs model-agnostic, there is no per-seat pricing, and you can deploy anywhere, from your own cloud to a fully air-gapped network. A new open model becomes one more option in the picker.
Each agent's LLM is chosen per agent, so a coding agent can try an open model while others stay where they are. Costs shows what each model spends, and Evaluations scores the outputs side by side.
ibl.ai is family-owned and operated from New York, NY, and our engineers work alongside your team to wire models like this into the systems you already run.
Sources: model specifications, license and benchmark tables from the Ripple-65B Preview model card; Beam's active parameters and our count of the 22-row table also from the card; founders, the $32 million figure (Ricursive's own, unaudited) and OpenRouter pricing from VentureBeat. The briefs framed Ripple as beating a model eight times its size; Beam is a mixture of experts with about 23B active parameters, and all benchmark figures are self-reported.