Ripple-65B Preview vs Ripple-65B: Why the Second Name Has No Model Behind It

Search for Ripple-65B Preview and you will find a documented model: a 65B dense Apache 2.0 checkpoint from Ricursive Intelligence, created 4 November 2026, with a card, four repositories and a vendor-reported 80.52 on Terminal-Bench 2.1. Search for Ripple-65B — the same name with the suffix trimmed — and you will find that model, over and over, under different filenames. There is no separate final release. The Hugging Face organisation lists four public repos — Ripple-65B-Preview, Ripple-65B-Preview-NVFP4, Ripple-65B-Base-Preview and Ripple-65B-Preview-DFlash2 — and not one of them drops the word "Preview."

That makes this an unusual comparison page, and the honest way to write it is to say so up front. Ripple-65B is not a rival to Ripple-65B Preview. It is a name people are searching for, a placeholder for a release that has not happened, and — most usefully — an occasion to answer the question those searches are really asking: are you getting the finished model, or are you getting a snapshot the lab told you was light on post-training?

What "Preview" means on this particular card

The suffix is not decorative here, and Ricursive is unusually explicit about it. The model card describes the release as relatively lightly post-trained: 200 billion mid-training tokens, 183 billion supervised fine-tuning tokens, 13.5 billion long-context SFT tokens, 700 RL steps, on top of a 12-trillion-token pretraining run at an 8K sequence length. Everything is trained from scratch, and the company's framing is that the post-training budget was deliberately small relative to the pretraining.

That combination — huge base, thin alignment — is what produces a checkpoint that scores 80.52 on Terminal-Bench 2.1 while sitting at 44.3 on DeepSWE 1.1 and 39.4 on SciCode. The agentic terminal performance is real work. The software-engineering breadth looks like the part that a longer post-training run would touch first. A final Ripple-65B would, on this reading, be a better model on exactly the axes where the Preview is weakest.

Ricursive has not committed to one. There is no roadmap, no "final release in Q4" line in the launch post, and no dated successor repo. When ArtificialWatch tracked the launch, it noted that "the GA without the Preview tag is still ahead" — an accurate summary of a state that has not changed since.

The four repositories, and what each one is actually for

The naming confusion gets worse before it gets better, because two of the four repos sound like they might be the final model and neither is.

• Ripple-65B-Preview — the primary instruct checkpoint. BF16, 64,782,689,360 parameters, 129.6 GB across 16 shards, Apache 2.0, ungated. This is the one every benchmark number refers to.

• Ripple-65B-Preview-NVFP4 — the same instruct model quantized with ModelOpt, 39,616,865,360 stored parameters, tagged 8-bit. Not a different model; a smaller container for the same weights.

• Ripple-65B-Base-Preview — created 14 October 2026, three weeks before the instruct release. This is the pretrained base, evaluated separately: MMLU 83.7, MMLU-Pro 61.5, GPQA Diamond 38.4, HumanEval 70.9, MBPP Sanitized 81.5. Those numbers look alarming next to the instruct card's 89.6 and 85.2 until you remember they are base-model scores — pre-SFT, pre-RL, answering in completion style. It is the correct starting point if you intend to do your own post-training, and the wrong one if you want an assistant.

• Ripple-65B-Preview-DFlash2 — 4.04B parameters, a speculative-decoding drafter. It proposes tokens for the 65B model to verify and cannot generate anything on its own; its card states outright that it "requires the target model to generate text."

Four repos, one model, three of them variants of a checkpoint and one of them a component. Nobody has published a fifth.

If a final Ripple-65B did ship, what would change?

This is the constructive part of the comparison, and it is worth being concrete rather than gesturing at "better post-training."

• The context story would probably get settled. The Preview advertises up to 1M, trained on 2M-token sequences in its long-context stages, while commercial listings have served 16,384 and, more recently, headlined 128K. A final release is the natural moment for a lab to say what the model is actually validated for rather than what the architecture supports.

• DeepSWE and SciCode would be the numbers to watch. Those are where the Preview trails models it beats elsewhere, and they are the benchmarks that respond most directly to more RL and more SFT on real repositories.

• An independent score would likely exist by then. Artificial Analysis has no entry for any Ripple slug today, which means every figure in circulation is vendor-reported. A GA release is the point at which aggregators usually pick a model up.

• The licence would probably stay Apache 2.0. Nothing in the launch suggests Ricursive intends to walk that back, and it is the single most valuable property the release has.

None of that is announced. It is inference from what the card says about its own post-training budget, and it should be read as such.

The naming trap, and why it matters practically

If you are wiring this model into a config file, the distinction stops being academic. The identifier that resolves is ricursiveAI/Ripple-65B-Preview. A request for ricursiveAI/Ripple-65B returns nothing, because nothing by that name exists — and a pipeline that silently falls back to a fuzzy match will either fail or, worse, quietly serve a base checkpoint where you expected an instruct one.

The same trap exists in reverse with the base model. Ripple-65B-Base-Preview is a real repository with real weights, and it is a genuinely different artifact from the instruct release: different scores, different behaviour, different purpose. Anyone comparing "Ripple-65B" scores across the internet is almost certainly mixing the instruct card's 89.6 MMLU with the base card's 83.7 without noticing they came from two different repos.

Three practical rules follow. Pin the exact repo id and a revision hash, not a family name. Do not assume a hosted endpoint and the BF16 download are the same product — the hosted one may be NVFP4, may be a different revision, and will almost certainly have a different context limit. And if a tool or a leaderboard lists "Ripple-65B" without the suffix, go and check which repository it means before you quote the number.

What you can do with the Preview today

Quite a lot, as it happens, and more than was true a week ago. The weights are ungated and Apache 2.0, so there is no approval step. As of 5 November, vLLM merged native support — a built-in RicursiveConfig, reasoning and tool-call parsers, and a recipe defaulting to NVFP4 on a single B300 — which means the checkpoint loads without --trust-remote-code in the engine most teams already run. A parallel SGLang pull request is still open, so vLLM is currently the only mainstream engine where this is a supported architecture rather than an external script.

Commercially, the model is served at roughly $0.20 per million input tokens and $0.60 per million output, with cached input billed at a fraction of that. Those rates describe the Preview, since it is the only thing there is to serve.

OrcaRouter does not carry it. That is worth stating plainly, because a comparison page that ends by implying you can reach both sides through us would be misleading — there is only one side, and we do not host it. What we do route is the surrounding field: Kimi K3.1 Pro at $2.00 and $10.00 per million with 1M context, DeepSeek V4 Pro at $0.66 and $1.98 with 1M context and 384K output, Qwen3.8 Max at $2.00 and $6.00. If the reason you are reading about a 65B open-weight coder is that you want a cheap agentic model without a procurement cycle, those are the ones on the same key as everything else, with failover and pass-through pricing. Ripple is the one you host or buy separately.

The short answer

If someone asks you the difference between Ripple-65B Preview and Ripple-65B, the accurate reply is that the second one does not exist yet. The Preview is the model: 65B dense, Apache 2.0, 12 trillion pretraining tokens, lightly post-trained, 80.52 on Terminal-Bench 2.1 by its own measurement, unranked by any independent board, and served at a context window that depends on who is serving it.

What a final release would add is a settled context figure, better numbers on the software-engineering benchmarks where the Preview is weakest, and probably the first outside verification the model has had. Until Ricursive ships it, every "Ripple-65B" comparison you read — including any that claim to compare it against the Preview — is comparing a model to a search term.