A lead of less than half a point
Start with the number Ricursive wants quoted. On Terminal-Bench 2.1, a test of agents completing real tasks in a terminal, Ripple scores 80.52 against 80.1 for Reflection Beam, 76.4 for Meta Muse Spark Open, 63.8 for Thinking Machines Inkling and 56.4 for NVIDIA Nemotron 3 Ultra. Every rival in that table is a mixture-of-experts model with 501B to 975B total parameters. Ripple has 65B.
Ripple is 65B parameters in total, compared with Beam's 501B, making Ripple roughly 7.7 times smaller by total parameter count. But the comparison changes when active parameters are considered. Beam activates 23B parameters per token, according to Ricursive's model card, while Ripple's dense architecture activates all 65B. That means Ripple uses roughly 2.8 times as many active parameters per token, based on the published figures. Total parameter count describes the size of the model; active parameter count is a more relevant, though still incomplete, indicator of per-token computational demand. Neither figure alone establishes actual inference speed or serving cost. Ripple is smaller by total parameter count but larger by active parameter count.
The margin itself deserves the same care. 80.5 against 80.1 is a lead, and at the top of a hard benchmark a lead is not nothing. It is also within the range where the choice of harness, context length and summarization moves scores by far more than 0.4. Ricursive's model card presents a separate comparison using the Terminus 2 harness at 128K. In that table, the company reports Ripple at 80.5, Beam at 70.8, Inkling at 56.2, Nemotron at 55.1 and Muse Spark Open at 78.7. These figures differ substantially from the launch blog's Terminal-Bench 2.1 comparison. The documents do not establish that the harness alone accounts for those differences, but they demonstrate how much the published rankings depend on the evaluation setup. A lead of four-tenths of one point cannot carry much weight when the maker's own second table moves the same rival by more than nine points.
Where the two tables part company
The harder problem is internal. Ricursive published two versions of its rival comparison on launch day, and they do not match.
On SWE-bench Pro v2.0 Hard, the launch graphic puts Muse Spark Open at 72.4 and Inkling at 39.6. The model card, in its own instruct-model table, puts Muse Spark Open at 52.3 and Inkling at 2.0. Ripple's 79.1 and Beam's 76.5 are the same in both, so the winner does not change. The size of the field behind it changes completely. A 2.0 and a 39.6 are not two estimates that can be reconciled through rounding. Both documents attach the figures to SWE-bench Pro v2.0 Hard, but neither explains why the rival scores differ so substantially. The discrepancy remains unresolved. Ripple's 79.1 and Beam's 76.5 match in both presentations, but that does not explain the differences elsewhere in the table.
DeepSWE 1.1 shows smaller versions of the same split, with Nemotron at 0.6 in the graphic and 0.2 in the card, and Inkling at 5.0 against 1.8. The card is candid about why this happens: it combines published scores with in-house evaluations where a lab had not published a result, it uses different harnesses for different benchmarks, and it states plainly that replicating modern agentic scores has been challenging because so many factors shift them. That candor is welcome. It also means the launch graphic, the version most likely to be screenshotted and shared, carries figures the company's own detailed table does not support.
Read in full, the card gives a more modest picture than the graphic. According to Ricursive's model-card evaluation table, Ripple leads three rows of the company's instruct-model comparison: both Terminal-Bench entries and SWE-bench Pro v2.0 Hard. In the same company-reported table, Ripple trails Beam on SWE-bench Verified and SWE-bench Pro v1.0, and trails Muse Spark Open on DeepSWE and SciCode. Its 39.4 on SciCode is the lowest score in that table, and it trails on most knowledge rows. That is a strong, specific result in terminal-based agentic coding. It is not a general coding crown. Ricursive's use of the qualifier "U.S. open-weight" also narrows its competitive framing. Reflection's own comparison places Chinese open-weight models above Ripple on Terminal-Bench, so the geographic qualifier matters when interpreting any claim about leadership.
The $32 million figure, and what it rests on
The second headline is cost. Ricursive says the whole effort cost $32 million, counting ladder experiments, data mixing, evaluations, debugging, synthetic data, posttraining development and the final runs, and estimates that at one-tenth of what the previous leader spent. Ricursive's one-tenth-cost estimate relies on a reported $150 million monthly compute commitment associated with Reflection and the reported four weeks of Beam pretraining plus four weeks of posttraining. This is Ricursive's calculation using reported figures, not a directly comparable accounting disclosure from both companies.
Ricursive reports that Ripple reached its Terminal-Bench result using 1.4 million reinforcement-learning rollouts. For comparison, Reflection has reported more than 100 million rollouts for Beam, as cited in coverage of Ricursive's announcement. These are company-reported figures, not independently audited counts, and they do not establish a comparable cost ratio: rollout lengths, evaluation expenses and computational requirements can differ substantially.
The supporting efficiency claims have the same limitations. Ricursive reports a 16-fold compute-efficiency improvement in its 4B-scale ladder experiments, measured on a single dolma_stack cross-entropy objective against the Olmo 3 ladder. Its model card additionally claims an advantage of up to 55-fold for the full hero run. These are company-reported training-efficiency comparisons, not measurements of coding-agent performance or inference costs. They provide evidence for Ricursive's training approach, but should not be translated into claims that Ripple is 16 or 55 times cheaper to serve.
None of this makes the cost story empty. A five-person team, by the card's account, trained a dense model from scratch at 40% model FLOPs utilization with simple sharding and no custom kernels, bought no posttraining data, and produced a model that belongs in the same evaluation conversation as systems built by far larger organizations. If an independent run confirms even part of the Terminal-Bench result, the efficiency argument becomes the most interesting part of this release.
What a team can actually do with it
The practical offer is concrete. The weights are downloadable now under a plain Apache 2.0 license with no revenue threshold, and as a base model. The BF16 weights would occupy approximately 130 GB, calculated from the 65B parameter count and two bytes per parameter, before overhead. A secondary measurement of the Hugging Face files put the NVFP4 checkpoint at approximately 57 GB; that size is not stated in the model card's text. Ricursive says its NVFP4 checkpoint can fit on consumer hardware such as the NVIDIA DGX Spark. The company also reports 250 tokens per second on two B200 GPUs at batch size eight, before quantization or custom kernels. On OpenRouter, checked November 5, the GMI Cloud endpoint lists $0.20 per million input tokens, $0.60 per million output tokens and $0.05 per million cache-read tokens.
There is a deployment detail worth checking before anyone budgets around the advertised context length. VentureBeat reported on November 4 that the OpenRouter endpoint was limited to 16,384 tokens. OpenRouter's model page, checked on November 5, lists a context window of 1,048,576 tokens. Those figures may describe different things: a provider-specific endpoint limit and the model-level context window. The listing does not, by itself, establish that every provider route supports the full advertised context. Teams should verify the actual limit on the endpoint they intend to use.
The test that matters is local anyway. Put Ripple on a team's own repositories, in its own agent harness, and record completed tasks, retries, token use and human interventions against the model it would replace. Ricursive has published enough detail, including the places where its own numbers conflict, for that evaluation to be worth an afternoon. It has not published enough for anyone to skip it.
Sources
- [1] Ricursive Intelligence (primary), "Introducing Ripple-65B Preview", launch blog, November 4, 2026 (read first-hand in full at Gate 1: Terminal-Bench 2.1 table — Ripple 80.52, Beam 80.1, Muse Spark Open 76.4, Inkling 63.8, Nemotron 3 Ultra 56.4; $32M full R&D definition and one-tenth estimate, with the footnote built from a reported $150M-per-month Reflection compute commitment and Beam's reported 4 weeks of pretraining plus 4 weeks of posttraining; 1.4M rollouts, roughly 70x fewer than Beam; 12T tokens, up to 1M context; dense architecture, 80 layers, d_model 5120, 40% MFU, no posttraining data bought; 16x ladder compute-efficiency claim at 4B scale on the dolma_stack cross-entropy objective against the Olmo 3 ladder; general-purpose variant wording; availability list)Read source
- [2] Ricursive Intelligence (primary), Ripple-65B-Preview model card, Hugging Face (read first-hand in full at Gate 1: Apache 2.0; BF16, NVFP4 and base checkpoints; active-parameter table — Ripple 65B active against Beam 23B active, Beam 501B total; Terminal-Bench 2.1 self-reported row and the separate Terminus 2 row — Ripple 80.5, Beam 70.8, Muse Spark Open 78.7, Inkling 56.2, Nemotron 55.1; SWE-bench Pro v2.0 Hard row with Muse Spark Open 52.3 and Inkling 2.0, Ripple 79.1 and Beam 76.5; DeepSWE row with Nemotron 0.2 and Inkling 1.8; SciCode 39.4; the disclosure that the table mixes official model-card scores with in-house runs for missing evals, uses Terminus 2 for SWE-bench Pro v2.0 Hard and DeepSWE 1.1 and OpenCode elsewhere, and warns that replicating agentic benchmark scores has been challenging; the up-to-55x hero-run efficiency claim; team scaled from 1 to 5 people; 250 tokens per second on 2x B200 at batch size 8 before quantization or custom kernels; DGX Spark positioning)Read source
- [3] VentureBeat, "Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks", November 4, 2026 (read first-hand in full at Gate 1: launch-graphic table with Muse Spark Open 72.4 and Inkling 39.6 on SWE-bench Pro v2.0 Hard and Nemotron 0.6 / Inkling 5.0 on DeepSWE 1.1; graphic-versus-card discrepancy reporting; Reflection's own published comparison listing several Chinese-developed models above Ripple's Terminal-Bench result; Reflection independently reported at more than 100 million rollouts, with the caveat that rollout counts do not by themselves establish a cost reduction; the one-tenth-cost estimate is not a directly comparable accounting disclosure from both companies; OpenRouter $0.20 / $0.60 pricing, $0.05 cached input, and the 16,384-token endpoint limit as checked November 4)Read source
- [4] OpenRouter, ricursiveAI/Ripple-65B-Preview listing and endpoints API (checked November 5, 2026 at Gate 1: $0.20 per million input tokens, $0.60 output, $0.05 cache read; FAQ lists a 1,048,576-token context window and up to 131,072 completion tokens; the GMI Cloud endpoint record in the endpoints API lists context_length 1,048,576; November 4, 2026 release date)Read source
- [5] MarkTechPost, "Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1", November 4, 2026 (secondary corroboration: three checkpoints under Apache 2.0; architecture and posttraining detail; Terminus 2 comparison with Beam at 70.8 and Muse Spark Open at 78.7; comparison table sourced to the model card and rival vendor sources, all self-reported)Read source