What Is Ripple-65B Preview? The 65B Open Model That Beats Bigger Rivals

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026 — a 65-billion-parameter dense open-weight model that, despite its size, posts the top self-reported agentic-coding score among US open-weight models. It beats models many times larger on key tests, ships under Apache 2.0, and fits on a single machine. Here is what Ripple-65B Preview actually is, how it benchmarks, and the caveats worth knowing before you trust the headline.

What is Ripple-65B Preview?

Ripple-65B Preview is a 65B-parameter dense Transformer, pretrained from scratch on 12 trillion tokens, with support for up to 1M tokens of context. It is released under the permissive Apache 2.0 license, with weights you can download and run yourself.

The model comes from Ricursive Intelligence, a Palo Alto startup better known for AI chip design than for frontier LLMs. Ripple is a general-purpose variant of the proprietary model that powers Ricursive's chip-design platform — now offered to everyone as an open-weight release. Co-founders Anna Goldie and Azalia Mirhoseini, who worked on Google's AlphaChip, lead the effort, and the launch drew public congratulations from Google's Jeff Dean.

The "Preview" label matters. By Ricursive's own account, the release went through relatively light post-training focused on agentic coding, so it is an early checkpoint rather than a finished product.

Ripple-65B Preview benchmarks

The headline claim is 80.52 on Terminal-Bench 2.1, a benchmark that measures how well a model writes and executes code inside a terminal. On Ricursive's model card, that score leads the US open-weight field it was compared against:

Two things stand out. First, every rival in that table is a Mixture-of-Experts (MoE) model with far more total parameters — Ripple is dense and much smaller, yet it edges them on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard. Second, the picture is not a clean sweep: Reflection Beam still leads on SWE-bench Verified, Muse Spark Open wins DeepSWE 1.1, and the larger models hold a slight edge on general knowledge (MMLU) and some math.

In other words, Ripple is strongest exactly where it was tuned — agentic coding — and merely competitive elsewhere.

Why a small dense model can beat bigger ones

Ricursive made a deliberate architecture bet. Instead of the MoE designs most labs now use, it trained a modernized dense Transformer to keep training and serving infrastructure simple. The model uses a "thin-and-deep" shape — a narrow d_model of 5,120 across 80 layers — and a Gemma-3N-like architecture chosen for training stability at scale.

The efficiency story is the real headline. Ricursive puts total R&D cost at roughly $32 million, which it estimates is about one-tenth the cost of the prior open-weight leader on agentic coding. It also reports reaching its Terminal-Bench score with 1.4 million RL rollouts — roughly 70x fewer than Reflection Beam used. Relative to its own Olmo 3 baseline, Ricursive reports a 16x compute-efficiency gain on code, rising to 55x at full scale.

Those numbers come from Ricursive, not an independent lab, so treat them as the vendor's self-reported figures until third parties reproduce them.

Can you run Ripple-65B Preview locally?

Yes — that is much of the point. At 65B total parameters, Ripple is small enough to fit on consumer-class hardware such as NVIDIA's DGX Spark, and Ricursive positions it for local development. It released three checkpoints: an instruction-tuned BF16 model, a quantized NVFP4 version, and a BF16 base model, plus a Colab to get started.

On datacenter GPUs, Ricursive reports around 250 tokens/second on 2xB200 at batch size 8, before any quantization or custom kernels.

You can get the weights and hosted access from:

• Hugging Face — ricursiveAI/Ripple-65B-Preview (Apache 2.0 weights)
• API partners — Together AI, GMI Cloud, and OpenRouter, with Fireworks listed as going live on launch day

How the "US open-weight SOTA" claim holds up

The claim is real but narrow. "State-of-the-art for US open-weight models" excludes closed models and non-US open models, so it is not a statement that Ripple is the best model, or even the best open model, overall. Independent coverage noted that Ripple trails some larger open models on other benchmarks.

A few honest caveats:

• Self-reported. The headline scores are Ricursive's own; Ricursive itself notes that agentic-eval scores shift a lot depending on the harness and settings used.
• Preview, not final. Light post-training means behavior may change in later releases.
• Scoped claim. The leadership claim is specific to US open-weight models on agentic coding, not a universal ranking.

None of that erases the achievement — a 65B dense model matching or beating 500B-to-975B rivals on hard coding tasks is a genuine result. It just means the framing deserves the same scrutiny as any launch-day benchmark.

Ripple-65B Preview vs Reflection Beam and Muse Spark Open

If you are choosing among recent open-weight coding models, the trade-off is mostly size versus breadth:

• Ripple-65B Preview — smallest and cheapest to run; best on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard; dense and easy to serve locally.
• Reflection Beam (501B) — stronger on SWE-bench Verified; much larger MoE, heavier to host.
• Muse Spark Open (816B) — wins DeepSWE 1.1 and MCP Atlas; a bigger Meta release with broad coverage.

For teams that want a locally deployable coding model and can accept "Preview" maturity, Ripple is the most interesting new option. For maximum breadth across coding evals, the bigger models still have an edge in places.

Run open models in your own AI workforce

Ripple-65B Preview is built for exactly the kind of work agents do all day: reading a repo, writing code, and executing commands in a terminal. Eigent is an open-source multi-agent "Cowork" desktop app that runs that workflow locally — point a team of agents at a task like reviewing GitHub pull requests and keep the model and your data on your own machine. If you want to pair a local open-weight model with a local agent runtime, download Eigent and start building your own AI workforce.