Ripple-65B Preview: A $32M Apache 2.0 Model Built to Fit a DGX Spark

What Ricursive released

Ricursive Intelligence published Ripple-65B Preview on 6 November, describing it as a general-purpose variant of the proprietary model behind its AI chip design platform. The model is a dense Transformer with 65 billion parameters, pretrained on 12 trillion tokens of mostly English text and code, and it supports up to 1,048,576 tokens of context.

The release includes an instruction-tuned model at full BF16 precision, an NVFP4 quantized version, and a pretrain-only base model, and the Hugging Face listing carries an Apache 2.0 license. That license is a real advantage over open models with revenue thresholds or branding conditions, because it leaves commercial and local use largely unrestricted.

Why 65B is the interesting number

Many recent open models chasing the top of the agentic coding charts are enormous. Ricursive's own comparison table lists Reflection's Beam at 501B, Meta's Muse Spark Open at 816B, and Thinking Machines' Inkling at 975B. Those are far outside what a single personal machine can reasonably hold.

Ripple goes the other way. The model card says that at 65B total parameters it "can fit on consumer hardware such as the NVIDIA DGX Spark", and Ricursive supplies the NVFP4 checkpoint and a notebook specifically to encourage local development. By our own arithmetic, 65 billion parameters at 4 bits each comes to about 32GB of weights before overhead, compared with about 130GB at BF16. That is an estimate, since neither the blog nor the model card we read states the NVFP4 file size.

Because every parameter is active on every token, 65B here is a different kind of number from the total counts quoted for mixture-of-experts models, which only activate a fraction of their weights at a time.

The benchmark lead is narrow and self-reported

Ricursive's headline claim is a US open-weight record of 80.52 on Terminal-Bench 2.1, just ahead of Beam at 80.1 and further ahead of Muse Spark Open at 76.4. A gap of less than half a point is not a decisive win, even before considering who ran the test.

Ripple's scores come from Ricursive, and some competitor numbers were filled in with its own in-house evaluation flow where official figures were missing. The company is candid about this, saying that replicating modern agentic evals is difficult and calling its results a best-effort comparison. BenchLM's listing repeats the model card's numbers and leaves Ripple without a computed overall score, so it does not act as an independent check. All of the published scores were also measured at full BF16 precision, which means they say nothing direct about how the 4-bit build performs.

The $32M claim and the from-scratch question

Ricursive says it developed and trained the model for $32M, covering its full research cost, and estimates that as a tenth of what the previous open-weight leader in agentic coding cost. The team grew from one person to five during development. The comparison is the company's own estimate, built partly on reported compute commitments, so treat it as indicative.

"From scratch" also needs a little care. The weights were pretrained from scratch, but the training recipe began from the Olmo 3 model ladder, which the team then improved. Crediting that open starting point is useful in itself, since it shows how public research recipes can lower the cost of entry for small labs.

What local AI users should wait for

Ricursive says the Preview had relatively light post-training with a focus on agentic coding, so general assistant behavior may lag its coding scores. The only speed figure on the card is 250 tokens per second on two B200 datacenter GPUs, and we found no published DGX Spark throughput or NVFP4 quality comparison.

Those are the two numbers to watch. If community testers confirm that the quantized model holds up at a usable speed, Ripple would be a rare combination: permissive licensing, a huge context window, and competitive coding on hardware one person can own.