Ripple-65B Preview: 80.52 On TerminalBench, Zero Independent Verification

Launch morning. GMI Cloud already had Ripple-65B Preview serving from its endpoints at $0.20 per million input tokens and $0.60 per million output tokens. And one test run pushed 100 concurrent coding agents through the model. All 100 fixes landed in 22.3 seconds, total cost about seven cents.

Cheerful version of the story, that one. Here is the rest.

Ricursive Intelligence, a company whose actual business is AI chip design, trained a 65-billion-parameter dense Transformer from scratch on 12 trillion tokens, posted the weights on Hugging Face under Apache 2.0.

And reported 80.52 on TerminalBench 2.1, presented as a new U.S. open-weight state of the art for agentic coding. The Decoder passed the number along with the tag "according to the company." Right tag. Every score in the launch table comes from Ricursive's own model card. Nobody outside the company has rerun it.

What Is Ripple-65B Preview?

Dense Transformer, 65 billion parameters, built by Ricursive Intelligence.

Trained from scratch. Not a finetune of some other lab's checkpoint. On 12 trillion tokens of predominantly English text and code, at 8K sequence length during pretraining. The model card describes it as a general-purpose variant of the proprietary model powering Ricursive's AI chip design platform, aimed at general-purpose, coding, reasoning, and agentic workloads.

Architecture choice is the unusual part.

Other labs went Mixture-of-Experts; Ricursive says it deliberately trained a modernized dense Transformer instead, codesigned to train and serve efficiently on NVIDIA hardware.

Dense at least means what it says on the tin.

Preview label matters too. The card admits "relatively light posttraining." Early checkpoint, then, not a finished product.

Ripple-65B Preview Specs: 65B Dense, 1M Context, Apache 2.0

The short version, for spec-lookup purposes:

| Spec | Value | |---|---| | Parameters | 65B, dense (no MoE) | | Pretraining corpus | 12 trillion tokens, predominantly English text and code | | Pretraining sequence length | 8K | | Context (model card) | up to 1Mi — 1,048,576 tokens | | License | Apache 2.0 | | Input / output | text / text |

One spec disagrees with itself. The vLLM serving recipe describes a 2M-token context, the model card says 1Mi, which is 1,048,576 tokens. And Fireworks lists up to 1M.

Working assumption: plan around 1M until the documents reconcile. Two first-party sources contradicting each other on the second-most-important spec says plenty about how young this release is. Long-context claims are cheap to print and slow to falsify, so a two-million-token monorepo in one prompt is not something to budget for yet.

Deeper in the recipe: 80 layers, alternating four 1,024-token sliding-window layers with one NoPE full-attention layer, plus a DFlash2 speculative drafter for faster serving. Three builds shipped — instruction-tuned BF16, instruction-tuned NVFP4, and a BF16 base model for anyone doing their own posttraining.

Ripple-65B Preview TerminalBench vs. 975B Models

The launch table, as published by Ricursive:

| Model | TerminalBench 2.1 (reported) | |---|---| | Ripple-65B Preview (65B, dense) | 80.52 | | Reflection Beam 501B | 80.1 | | Meta Muse Spark Open 816B | 76.4 | | Thinking Machines Inkling 975B | 63.8 | | NVIDIA Nemotron 3 Ultra 550B | 56.4 |

Two margins worth reading.

Over Reflection Beam 501B: 0.42 points, thin enough that nobody gets crowned. Over Thinking Machines Inkling 975B: nearly 17 points, against a model carrying 15 times the parameters. That second gap is the whole story if it survives contact with an outside evaluator.

Whether it survives is the open question.

One vendor chose the benchmark, quoted the competitors' scores on it, and published the table. Normal launch behavior. Also the exact recipe wrong leaderboards get made from.

The card rounds out the launch with 76.5 on SWE-bench Verified, 71.7 on MCP Atlas, 79.2 on IFBench, 98.3 on AIME 2025. And 85.2 on GPQA Diamond. Same provenance, every one of them.

Treat 80.52 as a claim, not a property.

Apache 2.0 Weights: Own, Don't Rent

License is the load-bearing fact here, and the benchmark chatter keeps burying it. Apache 2.0 on the weights means downloading them, running them wherever, modifying them. And shipping them inside a commercial product without a permission call to Ricursive. For a small operation, that is the line between renting a capability per token and owning it outright.

The design target lines up with the loop coding teams actually run all day: an agent plans a change, calls tools, reads the results, continues. Self-hosted, that loop keeps source on your own hardware instead of riding an API into somebody else's logs. For plenty of shops that is the entire argument.

Size is the counterweight.

Dense 65B in BF16 is a serious memory commitment, and no VRAM figure is published for any configuration.

The existence of an official NVFP4 build says quiet things about the expected deployment path.

Community artifacts are already moving. A GGUF conversion for llama.cpp, plus a repo targeting a single RTX PRO 6000 — but local means low precision, not BF16.

Hosting-averse route exists too. Fireworks AI and GMI Cloud both serve it, GMI at $0.20 per million input tokens and $0.60 per million output tokens from launch morning.

Ripple-65B Preview FAQ

Is Ripple-65B Preview really better than 975B models?

On the reported TerminalBench 2.1 table, it outscores Thinking Machines Inkling 975B 80.52 to 63.8 and NVIDIA Nemotron 3 Ultra 550B 80.52 to 56.4. Ricursive produced those numbers itself; no third party has repeated the run. Better on that benchmark, per its maker. Verified, no.

What license is Ripple-65B Preview under?

Apache 2.0. Commercial use, modification, and redistribution are permitted without asking anyone at Ricursive.

How much VRAM does Ripple-65B Preview need?

No published figure exists for any configuration. Dense 65B in BF16 is heavy, NVFP4 numerics cut the footprint, and the GGUF build opens llama.cpp as a route. Budget hardware before committing either way.

What context length does Ripple-65B Preview support?

The model card says up to 1Mi, which is 1,048,576 tokens. Fireworks lists up to 1M. The vLLM serving recipe says 2M. Plan around 1M until the sources agree.

Can Ripple-65B Preview run locally?

Weights are downloadable, a GGUF conversion for llama.cpp exists, and a community repo targets one RTX PRO 6000. The realistic local configuration is low precision, not BF16.

Worth an evening of testing, not a slot on the roadmap. Dense 65B, Apache 2.0, tool-calling focus. That combination matches what a self-hosting shop wants, and the 80.52 stays a claim until an outsider reproduces it. Both things hold at once. Pull the weights, wire in real tools, run five tasks off the actual backlog. And let those results overrule the launch post.

Sources

https://huggingface.co/ricursiveAI/Ripple-65B-Preview https://github.com/puretensor/Ripple-65B-Preview-1x-RTX-PRO-6000 https://fireworks.ai/models/fireworks/ripple-65b-preview https://www.gmicloud.ai/en/blog/we-tested-ripple-65b-preview-with-real-bugs-live-tools-and-planted-traps-here-is-what-held-up https://wiro.ai/models/ricursiveai/ripple-65b-preview https://huggingface.co/AtomicChat/Ripple-65B-Preview-GGUF https://benchlm.ai/models/ripple-65b-preview https://recipes.vllm.ai/ricursiveAI/Ripple-65B-Preview https://www.alphaxiv.org/ja/abs/2611.ripple https://the-decoder.com/ripple-65b-preview-delivers-strong-agentic-coding-results-as-a-fully-apache-2-0-open-us-model/