Ripple-65B Preview Claims Coding SOTA at $32 Million—Unverified

Ricursive Intelligence says it built a frontier-class open-weight model for $32 million—roughly one-tenth of what it estimates rivals spent—and the claim rests entirely on the company's own accounting.

The AI chip design startup published Ripple-65B Preview on November 4, 2026, under an Apache 2.0 license. It is a 65-billion-parameter dense Transformer, pretrained from scratch on 12 trillion tokens of mostly English and code data, with a 1-million-token context window. Ricursive describes it as a general-purpose variant of the proprietary model behind its chip design platform.

The $32 million figure covers pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, post-training recipe work, and the final hero runs, the company says. It arrives with a hedge: the number is Ricursive's own estimate, offered "to the best of our knowledge," and has not been independently audited. It is also unclear whether the tally includes opportunity costs or failed runs.

What do the benchmark claims actually show?

Ricursive self-reports 80.52 on Terminal-Bench 2.1, an agentic coding benchmark. It says that score beats far larger U.S. open-weight models: Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4). The company also reports 79.1 on SWE-bench Pro v2.0 Hard and 76.5 on SWE-bench Verified.

Benchmark aggregator BenchLM flags the Terminal-Bench 2.1 and SWE-bench Verified scores as "provider exact" and "display only"—self-reported, not independently verified. The SWE-bench Pro figure appears only in Ricursive's materials. BenchLM's best verified Terminal-Bench 2.1 score is 92.8% (SWE-2), leaving a 12.3-point gap above Ripple. The model trails elsewhere too: 39.4% on SciCode versus 60.1% for Sakana Fugu, 44.3% on DeepSWE versus 79.0% for Muse Wonder Full, and 31.3% on τ³-bench versus 96.0% for Mercury 2.5. Only 17 of 817 tracked benchmark slots have displayable evidence, and BenchLM excludes the model from public rankings pending more non-generated coverage.

The takeaway: the headline number is a company claim on a single benchmark family, not a verified result.

Why choose a dense model in 2026?

Ricursive skipped the mixture-of-experts designs common at this scale, arguing density reduces infrastructure complexity. The model uses a "thin-and-deep" layout—80 layers, a narrow d_model of 5,120—plus overprovisioned attention heads, a high sliding-window-to-global layer ratio, gated attention, global NoPE, value residuals, and FFN activation clipping.

The company reports holding 40% model FLOPs utilization through pretraining and post-training with plain sharding, no custom kernels, and no quantization. Pretraining followed a laddering approach starting from the Olmo 3 recipe; Ricursive says compounded changes delivered up to 16x compute efficiency on code at 4B scale. Post-training used no purchased data—only in-house synthetic pipelines and a mixed reinforcement learning recipe. Ricursive says it hit 80.5 on Terminal-Bench 2.1 with 1.4 million RL rollouts, roughly 70x fewer than Reflection Beam's reported 100 million.

On cost, the company estimates one-tenth of the prior open-weight state of the art, citing reports that Reflection committed $150 million per month for compute at SpaceX Colossus. That comparison depends on Reflection's reported figure, not audited accounts from either side.

How does it behave in the real world?

A hands-on test by MindStudio found the friction. A quantized build consumed about 75 GB of VRAM on a 96 GB machine with an NVIDIA RTX Pro 6000 Blackwell GPU. The supplied inference code had to be rewritten for the test hardware, and a custom serving layer with the Hermes agent framework was required before evaluation. Setup took several hours.

Once running, the model found a planted bug in a full-stack retirement calculator—code treating a 4% withdrawal rate as "4" instead of "0.04"—and produced a correct fix without guidance. It also solved a multi-table SQL/Databricks task involving nested arrays, window functions, and join logic.

API access has its own gap: OpenRouter currently offers a 16,384-token context window at $0.20/$0.60 per million input/output tokens, far below the documented 1 million-token maximum. Self-hosting at BF16 requires roughly 80 GB of VRAM—one H100 80 GB card.

Four checkpoints are available: a BF16 instruction-tuned model, an experimental NVFP4 quantized build targeting devices like the NVIDIA DGX Spark, a BF16 base model, and a DFlash2 drafter. Weights sit on Hugging Face; inference runs via GMI Cloud and OpenRouter, with Together AI listed as coming soon and Fireworks to follow.

What remains unknown?

No independent replication of any benchmark score exists. The proprietary chip-design variant has not been publicly evaluated. The company itself acknowledges light post-training and gaps in math, safety, and automation benchmarks, and further training rounds are planned. In plain terms, this is a promising preview, not a proven product.

Ricursive says it is hiring as it pushes toward "full-stack AI from model to silicon." Until independent evaluators verify the Terminal-Bench numbers, the market should treat Ripple-65B Preview as a cost claim with a benchmark attached, not a settled result.