There's another American victory worth paying attention to today: a new, open weights AI model that its creators claim scores the higher on coding performance on two third-party benchmarks compared to any other open model from the U.S.

It's called Ripple-65B Preview, and is the work of the lesser-known startup Ricursive Intelligence, which notes that it takes the lead on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard, two of the leading benchmarks used to measure performance at writing code in a terminal environment and difficult software engineering tasks, respectively (the latter just introduced back in September).

Ripple-65B Preview is available now for download on Hugging Face and via selected third-party application programming interface (API) partners including Together AI, GMI Cloud, and OpenRouter, with Fireworks support expected soon as well.

OpenRouter’s GMICloud endpoint charges $0.20 per million input tokens and $0.60 per million output tokens.

The launch comes with qualifications. Ricursive’s claimed lead applies to selected U.S. competitors and tests, and its own results show it trailing other open models on additional benchmarks.

The model’s advertised million-token capacity also exceeds the 16,384-token limit currently listed by OpenRouter.

A chip-design startup releases its underlying language-model technology

Ricursive’s announcement positions Ripple as an outgrowth of its semiconductor work. The company developed an internal language model to support its chip-design platform, then produced a public version for broader uses.

Ricursive described the release as “a general-purpose variant of the proprietary model that powers our AI chip design platform.”

That distinction matters: developers receive a related general-purpose model, rather than the company’s complete chip-design system.

The startup’s founders, CEO Anna Goldie and CTO Azalia Mirhoseini, previously helped create Google's reinforcement learning-based chip design software, AlphaChip.

Ricursive announced a $300 million Series A in January, led by Lightspeed at a $4 billion valuation, with participation from investors including Nvidia’s NVentures, DST Global and Sequoia.

Ripple therefore gives outside developers a way to assess technology from a well-funded startup whose main business centers on accelerating semiconductor development. Its coding benchmarks, however, do not establish the effectiveness of that separate chip-design platform.

Strong coding results, with a narrow lead on Terminal-Bench

Ripple contains 65 billion parameters, the numerical values a model learns during training. Ricursive trained it from scratch on 12 trillion tokens, units of text that can represent words, word fragments or punctuation.

The company emphasized tasks in which a model takes multiple steps and uses software tools to complete a job. Its launch materials reported the following results:

Benchmark

Ripple-65B Preview (65B)

Nvidia Nemotron 3 Ultra (550B)

Thinking Machines Inkling (975B)

Meta Muse Spark Open (816B)

Reflection Beam (501B)

Terminal-Bench 2.1

80.5

56.4

63.8

76.4

80.1

SWE-bench Pro v2.0 Hard

79.1

30.7

39.6

72.4

76.5

DeepSWE 1.1

44.3

0.6

5.0

60.0

43.3

HMMT February 2026

90.5

80.3

86.3

93.9

93.9

FORTRESS Adversarial

83.8

79.3

78.0

94.2

92.1

AutomationBench, public

36.4

8.4

14.6

42.4

39.3

On Terminal-Bench, Ricursive placed Ripple just ahead of Reflection’s Beam at 80.1, followed by Meta’s Muse Spark Open at 76.4, Thinking Machines’ Inkling at 63.8 and Nvidia’s Nemotron 3 Ultra at 56.4. Ripple’s advantage over Beam is approximately 0.4 percentage points using those rounded figures.

The company says it achieves the results with roughly 15 times fewer total parameters than Thinking Machines' Inkling.

The launch graphic also places Ripple first among that group on SWE-bench Pro v2.0 Hard. The pattern changes on DeepSWE, where Meta scores 60.0 against Ripple’s 44.3. Meta also leads Ripple on the displayed automation and adversarial safety tests, while Meta and Reflection both score higher on the math test.

A broader coding crown would overstate that evidence. Reflection’s own published comparison lists several Chinese-developed models above Ripple’s reported Terminal-Bench result.

Its updated Beam figures also differ from Ricursive’s comparison, including 77.2 on SWE-bench Pro v2-Hard and 44.4 on DeepSWE. Those are separate vendor evaluations, rather than a controlled test of both models.

Benchmark methods and conflicting figures complicate the comparison

There are discrepancies within Ricursive’s materials as well. Its model card lists SWE-bench Pro v2.0 Hard scores of 52.3 for Meta and 2.0 for Inkling, compared with 72.4 and 39.6 in its launch graphic. Some DeepSWE rival scores also differ.

The card says its tables combine published results with internal evaluations and acknowledges difficulty reproducing agent benchmarks. It also reports that changing the evaluation framework affected scores.

For buyers, this limits how confidently the figures can predict performance in an existing coding workflow. An agent’s surrounding software determines which tools it can access, how it manages a long conversation and how it handles failures. A benchmark result reflects that setup alongside the underlying model.

The reported results make Ripple a candidate for evaluation. They do not settle whether it will complete a company’s own coding tasks more accurately, quickly or cheaply than its current model.

How Ricursive says it reduced development costs

Ricursive also boasts of highly cost-efficient training for Ripple — with preliminary training experiments, dataset preparation, evaluations, troubleshooting, synthetic data production, development of its later training stages, and the final runs costing a total of $32 million, compared to the hundreds of millions or even billions estimated for other proprietary frontier models.

Its comparison with competitors deserves attribution. Ricursive estimates that it spent about one-tenth as much as the previous U.S. open-weight coding leader, using reported Reflection computing commitments and training durations to support the estimate. That is not a directly comparable accounting disclosure from both companies.

Two technical claims underpin Ricursive’s efficiency argument.

First, it said its reinforcement-learning process reached the reported Terminal-Bench result after 1.4 million rollouts. A rollout is a generated sequence of actions that researchers can evaluate and use as training feedback. Ricursive said this required roughly 70 times fewer rollouts than Reflection’s approach.

Reflection independently reported producing more than 100 million rollouts during a four-week reinforcement-learning run. The difference in counts does not by itself establish an equivalent cost reduction: action sequences can differ in length, hardware requirements and evaluation expense.

Second, Ricursive said it improved the efficiency of initial training by refining the model’s structure, training settings and data. Its experiments began with an Olmo 3 baseline and tested changes at smaller scales before committing to the larger model.

The company’s supplied training chart shows its final recipe reaching comparable code-prediction loss with up to 16 times less computation at the 4-billion-parameter scale. That is a measure of training efficiency on a specific objective. It does not translate directly into a 16-fold improvement in coding-agent performance or operating cost.

A dense model designed for simpler training and local deployment

Ricursive chose a dense architecture, which uses the model’s full parameter set during generation. It said that choice simplified its training infrastructure compared with a mixture-of-experts design, which routes work through selected portions of a larger model.

The company built Ripple with 80 layers and optimized its structure for Nvidia hardware. It said it maintained approximately 40% model FLOPs utilization—a measure of how effectively training uses available computing capacity—without specialized custom kernels or quantization during training.

The smaller total parameter count needs context. Ricursive’s model card lists Ripple at 65 billion active parameters, versus 23 billion for Beam, despite Beam’s 501 billion total. Total model size and computation per generated token are different measures.

For local deployment, Ricursive cites Nvidia’s DGX Spark, while its social announcement also names MacBooks. The release does not establish a single laptop memory configuration, speed or usable context length for that claim.

The model card links to both BF16 and compressed NVFP4 instruction-tuned weights, plus a BF16 base model. Developers evaluating local deployment will need to distinguish those versions and measure the resulting memory use and performance.

API pricing: a confirmed OpenRouter rate, with other providers still unresolved

The clearest published model-specific pricing is on OpenRouter, where GMICloud provides the endpoint.

Access route

Input, per million tokens

Output, per million tokens

Status checked November 4

OpenRouter, served by GMICloud

$0.20

$0.60

Public model and price listing

GMI Cloud directly

Not independently verified

Not independently verified

Named in Ricursive’s launch

Together AI

No Ripple-specific public rate located

No Ripple-specific public rate located

Model page lists forthcoming access

Fireworks AI

No confirmed Ripple-specific rate located

No confirmed Ripple-specific rate located

Ricursive announced availability later that day

Ricursive directly

No public rate card located

No public rate card located

Launch directs developers to outside providers

OpenRouter also lists cached input at $0.05 per million tokens. Its current endpoint supports 16,384 tokens of context, so developers should not assume this price purchases access to Ricursive’s advertised million-token window.

Together’s model page describes Ripple as coming soon to its serverless API. Its public pricing table did not include Ripple when checked.

Fireworks’ pricing documentation also lacked an individual Ripple entry. It publishes a general $0.90-per-million-token tier for unlisted dense models above 16 billion parameters, applying equally to input and output. That generic policy should not be treated as a confirmed Ripple launch price.

GMI Cloud’s public pricing page focuses on rented GPU capacity; its linked Ripple console page did not provide readable pricing during this review. The OpenRouter rate therefore establishes the cost of that particular access route, rather than a verified direct GMI quote.

Ripple’s listed rates are below Together’s published $0.40 input and $1.20 output prices for Beam. They do not make Ripple the least expensive hosted model: Together lists GPT-OSS-120B at $0.15 input and $0.60 output. These are comparisons across provider listings, without matching serving conditions.

For coding agents, token prices are only one part of the bill. Teams also need to measure how much text each task generates, how often the agent retries, and the cost of the computing environment in which it runs tools.

Licensing terms: Apache 2.0 permits commercial adaptation and redistribution

Ricursive’s repository contains an Apache 2.0 license, which permits commercial use, modification and redistribution subject to its conditions.

For enterprise developers, that allows adapting the model and incorporating it into commercial products without a model-license royalty. The published license contains no revenue or monthly-user threshold requiring a separate commercial agreement.

Redistribution still carries obligations. Recipients must receive the license, modified files must identify changes, and applicable copyright and attribution notices must remain. If the distribution includes a NOTICE file, relevant notices must accompany redistributed derivatives.

Apache 2.0 does not require developers to publish all their modifications under the same license. It also does not grant broad rights to Ricursive’s trademarks or provide a general warranty.

The weight license and a hosting provider’s service contract address different things. Downloading the model does not eliminate hardware and operating expenses, while using an API remains subject to that provider’s pricing and service terms.

What enterprise teams can evaluate now

Ripple’s immediate appeal is the combination of downloadable weights, a permissive license, competitive company-reported coding results and a low-priced hosted entry point.

The next step for a development team is a comparison using its actual repositories and agent software. That evaluation should record completed tasks, errors, time spent, token consumption and the amount of human intervention required. Teams interested in local use also need a separate assessment of memory requirements and performance.

Ricursive’s announcement makes a case for testing whether a substantially smaller total model can handle work otherwise assigned to larger U.S. open-weight systems. The remaining questions concern reproducibility, deployment limits and cost per completed job—the measures that determine whether a benchmark result becomes a useful enterprise tool.