Ripple-65B Preview: A New US Open-Weight Model, Explained

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026. It is a 65-billion-parameter dense Transformer, pretrained from scratch on 12 trillion tokens, with a context window of up to 1M tokens, published under Apache 2.0 on Hugging Face and served through Together AI, GMI Cloud and OpenRouter. Ricursive's claim is that it sets a US open-weight state of the art on Terminal-Bench 2.1 at 80.52, and that the whole project cost $32M. Both numbers are the vendor's own.

The company is not a model lab in the usual sense. Ricursive describes itself as a full-stack AI lab building self-improving systems across hardware, software and model, and says a proprietary variant of this model powers its chip-design platform. Ripple-65B Preview is the general-purpose release of that internal model. That origin explains two unusual choices: a dense architecture when most open-weight frontier models are mixtures of experts, and a published cost figure, which labs almost never give.

We read Ricursive's launch post, the Hugging Face model card and the OpenRouter listing on November 5, 2026. We found no independent benchmark of this model at that time, and we have not run one. Where this post gives a number, the next clause says who measured it.

Ricursive's launch post opens with the reason the model exists: to power its chip-design platform, it trained a language model from scratch, optimised the architecture for its own training and serving stack, and is now releasing a general-purpose variant. The Hugging Face card adds a detail the post leaves out, that the team scaled from one to five people during development. That is a small group for a 65B pretraining run, and it is part of why the cost claim in section 05 is worth examining rather than dismissing.

The release landed in four places on the same day. Hugging Face hosts the BF16 instruction-tuned model, an NVFP4 quantised checkpoint, a base model and a DFlash2 variant. Together AI, GMI Cloud and OpenRouter serve it through APIs, and the post said Fireworks would follow later that day. OpenRouter's catalogue record shows the route created on November 4, 2026, which for once matches the vendor's own date; a listing time is not a launch date, but here the two agree.

One example of what this is for: a team running a coding agent on an open-weight model today is usually choosing between very large mixtures of experts that need multi-GPU serving. Ripple's pitch is the same class of agentic coding score from a model one person can run on a desktop box. The failure case is reading "US open-weight state of the art" as "state of the art": the claim is scoped to US open-weight models on one benchmark, and Ricursive's own table shows rivals ahead on several others.

Almost every open-weight model Ripple is compared against is a mixture of experts: Reflection Beam has 501B total parameters and 23B active, Muse Spark Open 816B and 42B, Inkling 975B and 41B, Nemotron 3 Ultra 550B and 55B, all per Ricursive's table. Ripple has 65B total and 65B active. In plain terms, a mixture of experts stores far more knowledge than it uses on any one token, which lowers compute per token but requires all those parameters to sit in memory. A dense model uses everything it has, every time.

Ricursive gives an engineering reason for the choice rather than a quality one. The card says the aim was to reduce pretraining and post-training infrastructure complexity, and that the dense design let a small team train from scratch while holding 40% model FLOPs utilisation with simple sharding, no custom kernels and no quantisation. The launch post adds that a dense model gives strong reasoning with low-latency token generation, and that the small parameter envelope suits devices like the NVIDIA DGX Spark, which have high compute but low memory bandwidth relative to datacentre parts.

The trade-off the card does not hide is knowledge breadth. On MMLU, MMLU-Pro and GPQA Diamond, Ripple trails the leaders in Ricursive's own table by roughly three to five points, though it beats Reflection Beam on MMLU-Pro. The example that makes the design sensible is a coding agent that needs fast tokens and long context on one machine. The failure case is choosing it for a knowledge-heavy assistant because of the coding headline; the card's own numbers say the eight to fifteen times larger models know more.

Terminal-Bench 2.1 scores an agent on tasks completed in a terminal environment, and it is the benchmark Ricursive chose to lead with. The launch post states a US open-weight state of the art of 80.52, against Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4. The post also says the score was reached with 1.4 million reinforcement-learning rollouts, which it describes as roughly 70 times fewer than Reflection Beam used.

The model card then does something useful: it prints two Terminal-Bench rows. The first is self-reported scores, each lab's own. The second is Ricursive's re-run of every model through the official Terminus 2 harness at 128K context with summarisation. On that second row, Ripple stays at 80.5, Muse Spark Open rises to 78.7, Reflection Beam falls to 70.8, Inkling to 56.2 and Nemotron 3 Ultra to 55.1. Ricursive's own words are that replicating benchmark scores has been challenging, that scores were universally lower on the official harness except for Muse Spark Open, and that its results are a best-effort attempt at a fair comparison.

That footnote is the most honest thing in the release and the reason to hold the headline loosely. A 0.4-point lead over Beam on self-reported numbers becomes a 9.7-point lead on Ricursive's harness, and the gap to Muse Spark Open narrows from 4.1 to 1.8 points depending on which row you read. The failure case is quoting 80.52 as a settled ranking. The accurate statement is that Ricursive measured it, Ricursive explains why its rivals' published numbers may not be comparable, and nobody outside Ricursive has reproduced any of it yet. Our Terminal-Bench score guide covers why the harness matters as much as the model, and why a 2.1 score must never sit beside a 4.0 score.

Ricursive's comparison table runs to 22 rows across knowledge, maths, coding and agentic tasks, safety, instruction following and long context, and Ripple leads only a handful. It is first on Terminal-Bench 2.1 and on SWE-bench Pro v2.0 Hard at 79.1 against Reflection Beam's 76.5. It is second or lower on SWE-bench Verified at 76.5, where Beam has 81.6, on SWE-bench Pro v1.0 at 61.42 against Beam's 66.1, and on MCP Atlas at 71.67 where Muse Spark Open scores 89.2. On the maths rows it is within a few points of the leaders; on knowledge it trails all four on MMLU and GPQA Diamond, and beats only Reflection Beam on MMLU-Pro.

The card marks which rival scores are official model-card numbers and which come from Ricursive's in-house evaluation, and uses different harnesses for different tasks, Terminus 2 for the SWE-bench Pro Hard and DeepSWE rows and OpenCode elsewhere. That is more disclosure than most launch tables carry, and it is also a reminder that every cell is a vendor measurement under vendor conditions.

An example of how to read it: a team evaluating open-weight models for an agent that calls many tools should weight MCP Atlas and the τ³-bench banking row, where Ripple scores 71.67 and 31.27 against leaders at 89.2 and 36.1, more than the Terminal-Bench headline. The failure case is the opposite: treating one leading row as a summary of a 22-row table whose author took care to show where the model loses.

Labs rarely publish what a model cost. Ricursive did, and defined it. The $32M is described as the full R&D cost from scratch, including the development of the pretraining ladder, data mixing, evaluations, debugging, synthetic data generation, the post-training recipe, and the final pretraining and post-training runs. The company calls that one-tenth the cost of the previous open-weight leader on agentic coding, and attaches the words "as estimated to the best of our knowledge" with a footnote pointing to press reports of Reflection's compute commitments and Reflection's own account of Beam's eight-week training schedule.

Two things follow. The number is a vendor statement with no auditor, and the comparison rests on an estimate of a rival's spend that the rival has not published. The method claims behind it are more checkable in principle: a pretraining ladder derived from Olmo 3's, with the card reporting a 16x compute-efficiency gain on code at 4B scale and up to 55x for the full run, and a post-training recipe that bought no data and built its own synthetic pipelines. Those are the claims a researcher could try to reproduce; the dollar figure is not.

An example of the right use: a buyer deciding whether to trust a lab whose model team was five people at most can read $32M as a signal about the company's resources and intent, not as a quality metric. The failure case is turning it into a headline about the cost of frontier AI collapsing. One lab's self-reported bill for one preview model, compared against an estimate of someone else's, is not that.

OpenRouter's catalogue record for the route, retrieved November 5, 2026, lists $0.20 per million input tokens, $0.60 per million output tokens and $0.05 per million cached input tokens, with a 1,048,576-token context and 131,072 maximum completion tokens, served in BF16 with native tool calling. Those are route prices from the serving provider behind OpenRouter, not a price Ricursive has published; the launch post names no price of its own. Together AI and GMI Cloud list the model too, and we did not read their pricing pages for this post.

For local use, Ricursive publishes an NVFP4 checkpoint and a notebook, and names the NVIDIA DGX Spark as a target. A 65B model at four bits per weight, plus NVFP4's block scales, needs roughly 33 to 37 GB for the weights alone, before overhead and context, which is why a desktop box with large unified memory is the example the vendor reaches for. We have not run it on that hardware. Our DGX Spark comparison sets out what those machines can and cannot hold.

An example of the economics: an agent that consumes 50 million input tokens and produces 5 million output tokens a month would pay $13 at the OpenRouter route price. That is a reason to test it, not a reason to adopt it; the failure case is comparing a launch-week route price with a vendor's list price, when Ricursive has published none, and when launch-week routes on OpenRouter have changed within days before.

The licence is Apache 2.0, stated on the Hugging Face card with a link to the LICENSE file. That matters more than it sounds in a year when several open-weight releases shipped under bespoke licences with use restrictions. Apache 2.0 allows commercial use, modification and redistribution with attribution, and carries a patent grant. Our open-weight licence audit explains why the licence file, not the badge, is the thing to read, and that advice applies here too.

The preview label is doing real work. The card says post-training was relatively light and focused on agentic coding, the base checkpoint is pretrain-only with no mid-training or synthetic question-answer data, and the model was trained on a text-only corpus. There is no vision input, no published system card, no stated knowledge cutoff and no independent evaluation. Ricursive also says the pretraining data was kept clean, with minimal synthetic data, which is a choice that tends to lower traditional benchmark scores and which the base-model table reflects.

An example of a gap that matters to a buyer: there is no statement about data provenance beyond "our 12T text-only predominantly-English and code corpus", so a legal team cannot yet assess training-data risk from the release alone. The failure case is assuming a permissive licence answers that question. It answers what you may do with the weights, not where they came from.

The clearest fit is a team already running an open-weight coding agent that wants lower latency, a smaller serving footprint and a permissive licence, and is willing to run its own evaluation. The model's own card gives that team the benchmark rows to check first. The second fit is anyone who wants a serious 1M-context model on a single workstation; the NVFP4 checkpoint exists for that reader.

Teams that should wait are those whose workload is knowledge retrieval, multilingual, or multimodal, where the card's own numbers and scope statements point elsewhere, and those who need a system card or a third-party evaluation before procurement. For the open-weight field this model is entering, our Reflection Beam guide and Muse Spark Open guide cover the two models Ricursive measures itself against. Teams working out where an open-weight model belongs in their stack can bring that question to our AI transformation practice.