On November 4, Ricursive Intelligence released a set of weights that runs against the grain of the current industry. It is not another MoE scaled up to hundreds of billions of parameters, but a 65B dense Transformer with all 65B parameters active on every token. The official model card reports 80.52 on TerminalBench 2.1. The comparison table on that same card lists the models it is said to pass: 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 — models carrying roughly eight to fifteen times its total parameter count. It is worth being precise about where these numbers come from: the entire comparison is the Ricursive table, and the model card states that missing evaluations were filled in using an in-house eval flow.

65B beating 975B, and not on parameter count

What is really being challenged is not parameter count but route. Nearly every open-weight model today goes MoE: grow the total parameter count, use sparse activation to keep per-inference compute down, and accept that both model complexity and training infrastructure get harder. Ripple took the opposite route, and the stated reason is straightforward. A dense model buys low-latency token generation and much simpler engineering on both the training and serving side. The bet is plain: let every parameter participate, and get a model that is simpler to train and cheaper to serve.

The model was designed for the machine that runs it

The architecture was not picked at random. Ripple takes a thin-and-deep shape: a d_model of only 5120 spread across 80 layers, moving from Olmo 3 llama-like layout to a Gemma 3N-like one to keep training stable at scale. Attention heads are deliberately overprovisioned, while the sliding-window-to-global layer ratio goes up and the window size goes down to compensate for the cost. On top of that sit gated attention, global NoPE, value residuals, and FFN activation clipping. The result is a solid 40 percent MFU throughout pretraining and posttraining, achieved with simple sharding, no custom kernels, and no quantization — by a team that grew from one person to five. On the serving side, two B200s at batch size 8 reach 250 tokens per second before quantization, and the 65B model fits on an NVIDIA DGX Spark. A separate NVFP4 checkpoint and a getting-started colab are also published. The training recipe started from the Olmo 3 model ladder and was iterated up to v1.6, with a 2-stream mHC variant tried and dropped because it showed no gain at 4B scale. The full design rationale and the accepted and rejected trade-offs are in the launch blog.

Thirty-two million dollars and 1.4 million rollouts

The accounting is more revealing than the scores. For the entire effort — pretraining ladder experiments, data mixing, evaluations, synthetic data, posttraining recipe development, and the final training runs — the company says it spent 32 million dollars, roughly one tenth of the previous open-weight leader on agentic coding. That comparison is the company own estimate, benchmarked against a reportedly 150 million dollars a month compute commitment by Reflection. Posttraining reached 80.5 using only 1.4 million reinforcement learning rollouts, which the company puts at around one seventieth of what Reflection Beam used, the latter reportedly running past 100 million. The token budget per posttraining stage is also spelled out on the model card: 200B for midtraining, 6.5B for long context, 183B for SFT, 13.5B for long-context SFT, and 700 RL steps, with the long-context stages trained at a 2M sequence length. On the pretraining side, the recipe started from the Olmo 3 ladder and improved code compute efficiency by 16x at 4B model scale, reaching as much as 55x for the full hero run.

Discount the leaderboard

Re-run under the official Terminus 2 harness at 128K context and the rivals mostly fall: Reflection Beam drops from 80.1 to 70.8, Inkling from 63.8 to 56.2, and Nemotron 3 Ultra from 56.4 to 55.1, while Muse Spark Open actually rises to 78.7 — still below Ripple at 80.5. There are weak spots too. SciCode at 39.4 is the lowest of the five models, SWE-bench Verified at 76.5 trails both Beam at 81.6 and Inkling at 77.6, and MCP Atlas at 71.67 also sits behind. Chinese open-weight models are not part of the comparison at all, and IAMagazine checked that Ripple had not yet appeared in the Artificial Analysis Intelligence Index v5 at release time. No third-party independent replication exists yet.

So what

The interesting part is not yet another larger model. It is that the path of small-and-dense, runs-locally, and cost-auditable has been laid out in full for the first time: architecture, training recipe, hardware adaptation, and how much each step cost, all written down where anyone can check them line by line. TerminalBench-style agentic benchmarks are unusually sensitive to the evaluation harness, and the same set of models can shift as a whole when the harness changes, so that 80.52 deserves a discount. But if the approach holds, what is affected goes beyond a single terminal coding leaderboard to the premise that you need Mixture-of-Experts to reach the frontier at all.