A 65B Dense Model Beat Open Rivals Eight Times Its Size. The Cost Line Is the Real Story.

Ricursive Intelligence released Ripple-65B Preview on November 4, with weights on Hugging Face under Apache 2.0. The company reports 80.52 on Terminal-Bench 2.1, a test of an agent working inside a terminal. On its own comparison table that edges out Reflection's Beam 501B at 80.1, Meta's Muse Spark Open 816B at 76.4, Thinking Machines' 975B Inkling at 63.8, and NVIDIA's Nemotron 3 Ultra 550B at 56.4.
A 65B model sitting at the top of a table full of 500B-plus models is the headline. The line underneath it is more interesting: Ricursive says the whole thing cost $32 million, start to finish, and that it is roughly a tenth of what the previous record holder spent.
Dense, on purpose
Ripple is a dense Transformer with 64.8 billion parameters. Every parameter runs on every token. That is unusual in a year when the default answer to scale is a mixture of experts that activates a slice of its weights per token, and it is the single decision that makes the rest of the story possible.
A dense model asks less of your infrastructure. There is no router to balance, no expert parallelism to shard, no all-to-all communication pattern to tune. Ricursive says it trained at 40 percent hardware utilization with no custom kernels, starting from an Olmo 3 baseline rather than a blank sheet. The architecture is thin and deep: width of 5120 and 80 layers, with overprovisioned attention heads, gated attention and value residuals added to help reasoning.
The company's framing is that adapting the model and the training recipes to the hardware bought efficiency equivalent to training on 16 times more GPUs. Ricursive builds chip design tools, so co-design claims from it deserve a raised eyebrow. The measurable part is the utilization number and the absence of custom kernels, both of which are checkable against the training recipe if you have the appetite.
Context window is 1,048,576 tokens, pretraining was 12 trillion tokens of mostly English text and code, and the release includes a BF16 instruction-tuned checkpoint, an NVFP4 four-bit checkpoint, and a base model.
What $32 million covers
The launch post is explicit that $32 million includes pretraining experiments, data, evaluations, debugging and the final runs. That is the number people should argue about, because total program cost and cost of the last training job get conflated constantly.
Ricursive says the team grew from one person to five during development. It bought no post-training data and built its own synthetic pipelines. It reports 1.4 million rollouts to reach its Terminal-Bench score, roughly 70 times fewer than what Reflection's Beam needed. Fewer rollouts is the kind of saving that shows up in the bill and in the calendar.
If the numbers hold, the implication is uncomfortable for anyone whose strategy depends on capital as a moat. A five-person team, no custom kernels, one dense model, and a tenth of the budget producing a record on one benchmark. The counter is that a single benchmark is a narrow target and that the same team would need far more than $32 million to be competitive across the board.
Which brings us to the rest of the table.
Where Ripple loses
The model card is candid, and this is where it earns some trust. On SWE-bench Verified Ripple scores 76.5 against 81.6 for Reflection Beam. On SciCode it gets 39.4 against 58.5 for Muse Spark Open. On MCP Atlas, a tool-use test, 71.67 against 89.2. On DeepSWE 1.1, 44.3 against 60.0.
So the pattern is a model that is excellent at driving a terminal and mediocre at scientific code and tool orchestration relative to the biggest open competitors. MMLU lands at 89.6 and AIME 2024 at 97.1, both slightly behind the largest models, which is what you would expect from a 65B against an 816B.
The card also admits something most cards skip: replicating agentic benchmark scores is hard, and some cells in the table use Ricursive's own runs where the other lab had not published a number. It notes that Terminal-Bench scores came out lower for nearly every model when run with the official Terminus 2 harness at 128K sequence length. Read that sentence twice. It means the comparison table is internally consistent and not necessarily portable.
The record may also have lasted about a week. Thinking Machines released Inkling 2 Small on November 5, and its model card reportedly puts Terminal-Bench 2.1 at 81.3 percent on the same Terminus harness. Those weights stayed closed at launch, so it does not count as an open-weight record, and the lab said Apache 2.0 weights would follow after external safety testing.
The local claim, unpacked
Ricursive says Ripple fits on an NVIDIA DGX Spark and even on a MacBook. Neither post says how much memory that takes, and the full-precision files on Hugging Face add up to around 130 GB, which is well beyond a laptop.
What makes local use plausible is the quantisation work. The repository ships an NVFP4 checkpoint and a Colab notebook, and the Hugging Face page lists around 30 quantised versions already. At two B200s with batch size 8 the model generates 250 tokens per second before quantisation or kernel tuning, which is a reasonable figure for a dense 65B.
The appeal for a law firm or a clinic is obvious: prompts and code never leave the network, and Apache 2.0 allows commercial use without paying Ricursive a licence fee. The catch is that this is a Preview with light post-training aimed at coding, and the repository ships custom code in `modeling_ricursive.py`, which means trust-remote-code style loading. A platform team should read that file before running it anywhere that matters.
Why a chip company is shipping a model
Ricursive's business is chip design tooling, and Ripple is a general-purpose version of the model behind that platform. The company's argument is that model architecture and chip architecture have to be designed together, and shipping an open model is a way to prove the claim in public.
That is a different motivation from the frontier labs, and it changes what to expect. Ricursive does not need to win on breadth. It needs to show that a small, well-matched model on well-matched hardware can do serious agentic work, because that is the pitch it sells to customers who cannot buy a data center.
Whether the benchmark record survives scrutiny matters less than whether the recipe does. A dense 65B trained for $32 million, running on a desktop box, under a permissive licence, is a specific and testable proposition. If it holds, the interesting question stops being who has the biggest model and becomes who can afford to keep pretending size is the only axis.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.