Ricursive Intelligence Releases Ripple-65B Preview, an Apache 2.0 Dense Model That Reports a Terminal-Bench 2.1 Lead Over Selected U.S. Open Models
Ricursive Intelligence released Ripple-65B Preview, a 65B dense model listed under Apache 2.0, reporting 80.52 on Terminal-Bench 2.1; VentureBeat notes mixed results elsewhere and conflicting figures.
Overview
Ricursive Intelligence has released Ripple-65B Preview, a dense Transformer with 65 billion parameters, and the model card on Hugging Face lists its license as Apache 2.0. The company reports a Terminal-Bench 2.1 score of 80.52 for the model, ahead of four other U.S. open-weight models it compares against. VentureBeat reports that the claim “applies to selected U.S. competitors and tests” and that the company’s own tables show the model trailing on other benchmarks.
What was released
According to the model card, the Preview release includes an instruction-tuned model in BF16 and NVFP4 numerics and a BF16 base model. The card describes a model pretrained on 12 trillion tokens with support for up to 1Mi (1,048,576 tokens) of context, a d_model of 5120 and 80 total layers. It says the release “has gone through relatively light posttraining, with a focus on agentic coding.” The LICENSE file in the repository begins with the heading Apache License, Version 2.0, January 2004.
VentureBeat quotes Ricursive as describing the model as “a general-purpose variant of the proprietary model that powers our AI chip design platform.” It writes that CEO Anna Goldie and CTO Azalia Mirhoseini previously helped create Google’s reinforcement learning-based chip design software, AlphaChip, and that Ricursive announced a $300 million Series A in January, led by Lightspeed at a $4 billion valuation.
Access and pricing
VentureBeat reports that the weights are on Hugging Face and that API partners include Together AI, GMI Cloud and OpenRouter, with Fireworks support expected soon. As of the November 4 check described in its article, OpenRouter’s GMICloud endpoint charged $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05 per million tokens. VentureBeat adds that the endpoint supported 16,384 tokens of context, against the million-token capacity Ricursive advertises, and that it located no Ripple-specific public rate for Together AI or Fireworks AI at that time. Prices and availability may have changed since.
The benchmark claims
The model card says Ripple-65B Preview scores 80.52 on TerminalBench 2.1, ahead of 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). In the card’s comparison table, Ripple-65B Preview also scores 79.1 on SWE-bench Pro v2.0 Hard against 76.5 for Reflection Beam 501B. VentureBeat puts the Terminal-Bench margin over Beam at approximately 0.4 percentage points using rounded figures.
The same table lists Ripple-65B Preview behind at least one of the four comparison models on several other tests. On DeepSWE 1.1 it scores 44.3 against 60.0 for Meta Muse Spark Open 816B, and on SWE-bench Verified it scores 76.5 against 81.6 for Reflection Beam 501B. On GPQA Diamond it scores 85.2, with the four comparison models between 87.0 and 90.6.
The card says that “Replicating benchmark scores has been challenging” and calls its results “a best-effort attempt at a fair comparison.” It states that scores for missing evaluations come from Ricursive’s own in-house evaluation flow, and that it used the Terminus 2 harness for Ripple-65B Preview on SWE-bench Pro v2.0 Hard and DeepSWE 1.1 and OpenCode in other cases where no harness was specified.
Caveats VentureBeat raises
VentureBeat writes that “A broader coding crown would overstate that evidence.” It reports that Reflection’s own published comparison lists several Chinese-developed models above Ripple’s reported Terminal-Bench result, and that Reflection’s updated Beam figures include 77.2 on SWE-bench Pro v2-Hard and 44.4 on DeepSWE. It describes these as separate vendor evaluations rather than a controlled test of both models.
VentureBeat also reports discrepancies within Ricursive’s materials: it says the model card listed SWE-bench Pro v2.0 Hard scores of 52.3 for Meta and 2.0 for Inkling, against 72.4 and 39.6 in the launch graphic. The card text read for this article on November 6 lists 72.4 and 39.6 in that row, so the card may have changed after VentureBeat’s review; the Herald did not see the earlier version.
In both cited sources, the Ripple-65B Preview benchmark figures are supplied by Ricursive. VentureBeat introduces its results table as what the company’s launch materials reported, and the model card is the company’s own document. Neither cited source describes an evaluation of Ripple-65B Preview run by a party independent of Ricursive.
Cost and efficiency claims
VentureBeat reports that Ricursive puts total development cost at $32 million, covering preliminary experiments, dataset preparation, evaluations, synthetic data and the final runs, and that Ricursive says it reached the reported Terminal-Bench result after 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection’s approach. VentureBeat notes that Reflection reported more than 100 million rollouts during a four-week run, and says Ricursive’s one-tenth-the-cost comparison “is not a directly comparable accounting disclosure from both companies.”
On hardware, the model card says that at 65B total parameters the model can fit on consumer hardware such as the NVIDIA DGX Spark, and that it reaches 250 tok/s on 2xB200 batch=8 prior to quantization or custom kernels. These are Ricursive’s figures. The card also lists Ripple at 65B active parameters, versus 23B for Beam, which VentureBeat says means total model size and computation per generated token are different measures.
Licensing
Earlier Herald coverage described Kimi K3.1 Pro weights published under a license requiring a separate agreement for model-as-a-service businesses above $20 million, and Mistral Large 4 open weights under a modified MIT license that requires a commercial agreement for such businesses. For Ripple, VentureBeat writes that the published license contains “no revenue or monthly-user threshold requiring a separate commercial agreement.” The Herald has not independently reviewed the full license text for obligations such as notices or attribution.
What remains open
VentureBeat’s own conclusion is that the results make Ripple a candidate for evaluation and “do not settle” whether it will complete a company’s own coding tasks more accurately, quickly or cheaply than a current model. It recommends comparing models on a team’s own repositories and agent software, recording completed tasks, errors, time, token consumption and human intervention.