Ripple-65B Preview · Ricursive · November 2026

Ripple-65B Preview: A 65B Open Model With a Headline Score That Needs Two Footnotes

On November 4, 2026, Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense open-weight model for agentic coding, under the Apache 2.0 license. Its headline — 80.5 on Terminal-Bench 2.1, ahead of far larger US open models — is real as a vendor-reported number, and it is not yet a verified verdict. This page keeps the two things apart: what Ricursive actually shipped, and the two checks that change how the headline reads.

Last updated: November 5, 2026. Facts checked against Ricursive's launch materials as reported by VentureBeat, MarktechPost, StartupHub, and CryptoBriefing on November 4–5, 2026, plus an independent comparison of the Terminal-Bench versions. All benchmark figures below are vendor-reported unless stated otherwise; no independent reproduction had appeared as of this page's date, and this page will be updated when one does.

At a Glance

What it isRipple-65B Preview — a dense, open-weight model for agentic coding, a general-purpose variant of the proprietary model behind Ricursive's AI chip design platform
ReleasedNovember 4, 2026, by Ricursive Intelligence; Apache 2.0 license; three checkpoints: instruction-tuned BF16, NVFP4, and BF16 base
Size / training65B dense (all parameters active per token), 80 layers, 5,120-wide residual stream, pretrained from scratch on 12 trillion tokens, up to 1M context claimed
Headline score80.5 on Terminal-Bench 2.1 (vendor-reported, not independently reproduced); 79.1 on SWE-bench Pro v2.0 Hard (vendor-reported)
Where to try itHugging Face weights; API endpoints at Together AI, GMI Cloud, and OpenRouter; Fireworks expected later
Reported priceAround $0.20 input / $0.60 output per million tokens at hosted endpoints (reported, not a vendor list price — confirm on the provider page)
Hardware barBF16 build is a datacenter-GPU model (vendor-reported 250 tokens/s on 2x B200); experimental NVFP4 build targets NVIDIA DGX Spark-class hardware

Check 1: 80.5 on Terminal-Bench 2.1 Is Not 66.4 on Terminal-Bench 4.0

The comparison doing the rounds puts Ripple's 80.5 next to Claude Opus 5.5's 66.4 and reads it as a fourteen-point win for the open model. That reading is wrong, because the two numbers come from different exams. Terminal-Bench 4.0 removed eight tasks, fixed nineteen more, applied a flat eight-hour timeout, and added tamper-resistance checks. When the same models were run across both versions, scores fell sharply on 4.0 — one model dropped about seventy points without changing at all. Until Ripple is run on Terminal-Bench 4.0, there is no honest head-to-head number against Opus 5.5 on agentic coding.

The same-version comparison that does exist points the other way. On DeepSWE 1.1, where neither lab has re-versioned the test, Ricursive reports 44.3 for Ripple and Anthropic reports 74.2 for Opus 5.5. That single row does not settle anything either — it is one benchmark, and Ripple's strength is terminal-based coding tasks — but it is the reason this page does not present 80.5 as proof that a 65B open model beats a frontier closed model. The honest summary is narrower: on Terminal-Bench 2.1, Ripple leads the US open-weight field Ricursive chose to compare against, by less than half a point over Reflection Beam.

Check 2: The Vendor Table Itself Is Messy

Even inside Ricursive's own materials, the table needs care. VentureBeat found discrepancies between the launch graphic and the model card on SWE-bench Pro competitor scores — for example, different figures for Meta and Thinking Machines models depending on which document you read. Ricursive's card also notes difficulty reproducing agent benchmarks and that its tables mix published results with internal evaluations. None of this means the headline is false; it means the table is a vendor presentation, not a neutral leaderboard, and competitor rows should not be quoted as independently verified.

Ricursive's own table is also candid about where Ripple is weak: it trails larger models on knowledge and science coding tests, with SciCode at 39.4 the lowest of the five models compared. That pattern — strong on terminal coding tasks, weaker elsewhere — fits a model trained specifically for agentic coding, and it is a better buying signal than the headline alone: if your work is terminal-based coding agents, Ripple is worth a trial; if it is broad knowledge work, the table itself argues against picking it on this evidence.

What You Actually Get, and What It Costs to Run

The release includes three checkpoints under Apache 2.0: an instruction-tuned BF16 model, an NVFP4 version, and a BF16 base model. Weights are on Hugging Face, and hosted endpoints are listed at Together AI, GMI Cloud, and OpenRouter, with Fireworks expected later. Reported endpoint pricing is around $0.20 per million input tokens and $0.60 per million output tokens — cheap for a coding model, but a reported rate, not a vendor list price, and the advertised 1-million-token context exceeds the 16,384-token limit then listed by OpenRouter, so check the context window on the endpoint you actually use before building a long-context workflow on it.

Self-hosting is a datacenter proposition. The BF16 build is reported at 250 tokens per second on two B200 GPUs, and the experimental NVFP4 build targets NVIDIA DGX Spark-class hardware, which Ricursive notes has high FLOPs but low memory bandwidth versus datacenter accelerators. A dense 65B model activates every parameter on every token, so do not compare its serving cost to a mixture-of-experts model with a similar total parameter count but far fewer active parameters. For most developers, the realistic path is a hosted endpoint, not a local run.

Why Ricursive Thinks a Smaller Dense Model Wins

Ricursive's development story is the most interesting part of the release, and it is also entirely vendor-reported. The team says it used model laddering — systematic small-scale experiments to pick architecture and data mixes before a single large run, starting from the Allen Institute for AI's Olmo 3 ladder — and reached its Terminal-Bench 2.1 score with 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection Beam. Total R&D cost is put at $32 million end to end, which Ricursive estimates at about one-tenth of the previous open-weight state-of-the-art cost on agentic coding. If those efficiency claims hold up under independent scrutiny, they matter more than the 0.4-point lead over Beam: they suggest open-weight coding models can be built far more cheaply than assumed. Until then, they are claims to verify, not facts to quote.

Related guides on this site: Mistral Large 4 — a large open-weight model guide, DeepSeek Super-agent — a Router-Worker coding prototype, Cindy — a multi-harness agent that switches tools mid-task, and GPT-6.1 Terra — OpenAI's small coding model.

Sources checked November 5, 2026: VentureBeat's launch coverage, MarktechPost's model breakdown, StartupHub's and CryptoBriefing's launch reports, and an independent comparison of Terminal-Bench 2.1 versus 4.0 scoring. Where those sources disagree with each other or with Ricursive's model card — notably on SWE-bench Pro competitor rows — this page says so instead of picking one figure silently. Independent reproduction status: unconfirmed as of November 5, 2026.

FAQ

Is Ripple-65B Preview really better than Claude Opus 5.5 at agentic coding?

Not on the evidence published so far. Ricursive reports 80.5 on Terminal-Bench 2.1, while the often-quoted 66.4 for Claude Opus 5.5 is on Terminal-Bench 4.0 — a different, reworked test set with removed and fixed tasks, a flat eight-hour timeout, and tamper-resistance checks. Putting the two numbers in one column is misleading. On DeepSWE 1.1, where both labs published a same-version score, Ricursive reports 44.3 for Ripple and Anthropic reports 74.2 for Opus 5.5. No independent reproduction of Ripple's headline scores had appeared as of November 5, 2026.

What is Ripple-65B Preview?

Ripple-65B Preview is a 65-billion-parameter dense open-weight model released on November 4, 2026 by Ricursive Intelligence, a startup whose proprietary model powers its AI chip design platform; Ripple is described as a general-purpose variant of that model. Dense means all 65 billion parameters are active for every token, unlike mixture-of-experts models that activate only part of the network. It is pretrained from scratch on 12 trillion tokens, uses an 80-layer thin-and-deep Transformer with a 5,120-wide residual stream, claims up to 1 million tokens of context, and is released under the Apache 2.0 license with three checkpoints: an instruction-tuned BF16 model, an NVFP4 version, and a BF16 base model.

What do Ricursive's own benchmark numbers actually say?

Ricursive's self-reported table puts Ripple at 80.5 on Terminal-Bench 2.1, ahead of 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 — a margin of less than half a point over Beam. It also reports 79.1 on SWE-bench Pro v2.0 Hard. The same materials show Ripple trailing larger models on other knowledge and reasoning tests, with SciCode at 39.4 the lowest of the five models in its own table. VentureBeat also flagged discrepancies inside Ricursive's materials on SWE-bench Pro competitor scores between the launch graphic and the model card, so treat competitor rows as vendor-presented, not independently verified.

Where can I try Ripple-65B Preview, and what does it cost?

The weights are available on Hugging Face, and hosted API endpoints are listed at Together AI, GMI Cloud, and OpenRouter, with Fireworks support expected later. Hosted endpoint pricing is reported at around $0.20 per million input tokens and $0.60 per million output tokens; there is no vendor list price to quote, so check the provider's pricing page before buying. One caveat reported at launch: the advertised 1-million-token context exceeds the 16,384-token limit then listed by OpenRouter, so confirm the context window on the endpoint you actually use.

Can I run Ripple-65B Preview on my own machine?

For most people, no — not in full precision. The BF16 build is a datacenter-GPU model, with vendor-reported throughput of 250 tokens per second on two B200 GPUs. An experimental NVFP4 build targets NVIDIA DGX Spark-class hardware. A 65B dense model activates every parameter on every token, so it is far heavier to serve than a mixture-of-experts model of similar total size. Unless you have datacenter GPUs or DGX Spark-class hardware, use a hosted endpoint instead of self-hosting.

Why did Ricursive build a 65B dense model instead of a bigger MoE?

Ricursive's stated argument is efficiency of development, not just size: it says the model reached its Terminal-Bench 2.1 score with 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection Beam, and that pretraining plus post-training cost $32 million end to end — which it estimates at about one-tenth of the previous open-weight state-of-the-art cost on agentic coding. Both figures are Ricursive's own claims, built on a model-laddering method adapted from the Allen Institute for AI's Olmo 3 ladder. They describe how the model was developed, not proof that it will beat larger models on your tasks.

Turn ideas into AI video with HAILUOAI

Explore HAILUOAI's text-to-video and image-to-video tools to create your own AI-generated videos.