Chip-design startup releases its first model
U.S. AI startup Ricursive Intelligence released the weights (trained parameters) of "Ripple-65B Preview," its first model, on Nov. 4 (Nov. 5 Japan time). It has 65 billion parameters, and the company handled everything from pretraining onward itself. The company claims the model has reached the top level among U.S. models with released weights in "agentic coding," in which AI carries out steps on its own to build programs. It said the model outperformed far larger models such as "Inkling" (975 billion) and "Nemotron 3 Ultra" (550 billion). All figures are the company's own, and there has been no third-party verification yet.
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch. Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Ricursive was founded by CEO Anna Goldie and CTO Azalia Mirhoseini, who developed Google's semiconductor design AI "AlphaChip." The company said the new model is a general-purpose version of a private model that supports its AI chip design platform. In January 2026, it raised $300 million in a round led by Lightspeed Venture Partners, at a valuation of $4 billion. Nvidia's investment arm NVentures and Sequoia Capital are also investors.
In the United States in October, Meta's "Muse Spark Open" and startup Reflection AI's "Beam" released their weights one after the other. According to DeepLearning.AI, Chinese models dominate the top of evaluations of open models.
80.52 on Terminal-Bench 2.1, 0.42 points ahead of Beam
The basis for the top-level claim is the score (out of 100) on Terminal-Bench 2.1, which tests solving tasks through computer terminal operations. Ripple-65B Preview scored 80.52, 0.42 points ahead of Beam's 80.1. Muse Spark Open scored 76.4, Inkling 63.8 and Nemotron 3 Ultra 56.4.
In results Ricursive remeasured using the same evaluation harness, "Terminus 2," for all models, Ripple-65B Preview stayed at 80.5, Muse Spark Open scored 78.7 and Beam 70.8, swapping second and third places. All competitor figures are from the company's internal measurements.
The comparison of model size comes with a caveat. The four rival models are "mixture of experts (MoE)" types that use only some components for each step, activating 23 billion to 55 billion parameters at a time. Ripple-65B Preview is a dense model that runs all 65 billion parameters every time. Its total parameter count is smaller than the rivals' 501 billion to 975 billion, but it activates more at a time than the rivals do.
Fourth of five on SWE-bench Verified
In other evaluations listed in the model card, it falls short of rivals on some benchmarks. On SWE-bench Verified, which tests fixing real software bugs, it scored 76.5, short of Beam's 81.6, Muse Spark Open's 78.7 and Inkling's 77.6. It topped Nemotron 3 Ultra's 71.9.
On MCP Atlas, which measures the ability to use external tools, it scored 71.67, below Muse Spark Open's 89.2, Beam's 80.8 and Inkling's 76.0. It scored 39.4 on SciCode, which involves writing scientific computing programs, and 85.2 on GPQA Diamond, which tests specialized knowledge, the lowest of the five models on both. Inkling, which Ricursive claims to have outperformed, also scores higher than Ripple-65B Preview on these three evaluations and on SWE-bench Verified. Nemotron 3 Ultra also comes out ahead with 44.6 on SciCode and 87.0 on GPQA Diamond.
Many comparison figures are internal measurements; third-party evaluations still to come
There are also benchmarks on which Ripple-65B Preview ranks first. On SWE-bench Pro v2.0 Hard, a set of difficult development tasks, it scored 79.1, ahead of Beam's 76.5 and Muse Spark Open's 52.3. However, Muse Spark Open's 52.3, Nemotron 3 Ultra's 30.7 and Inkling's 2.0 are Ricursive's internal measurements. On DeepSWE 1.1, Nemotron 3 Ultra's 0.2 and Inkling's 1.8 are also internal measurements. On these two evaluations, Ripple-65B Preview was measured with Terminus 2, which differs from OpenCode, the harness used for the other evaluations.
No third-party comparison was found as far as ORIENS checked. ORIENS could not confirm a listing on Artificial Analysis' overall index as of Nov. 5. Ricursive's comparison table lists only U.S. models and does not include Chinese models. On the same day, Nov. 4, China's Moonshot AI released the weights of "Kimi K3.1 Pro," which scored 46 on Artificial Analysis' overall index (version 5). DeepLearning.AI reported on Oct. 23 that on the pre-revision version of that index, Muse Spark Open was the top U.S. open model but trailed China's GLM-5.3 and others.
| Benchmark | Ripple-65B Preview | Nemotron 3 Ultra | Inkling | Muse Spark Open | Beam |
|---|---|---|---|---|---|
| SWE-bench Pro v2.0 Hard | 79.1 | 30.7* | 2.0* | 52.3* | 76.5 |
| DeepSWE 1.1 | 44.3 | 0.2* | 1.8* | 60.0 | 43.3 |
| MCP Atlas | 71.67 | 62.2* | 76.0 | 89.2 | 80.8 |
| AA-LCR | 75.0 | 65.4 | 78.3* | 80.7* | 78.2 |
Development cost put at about $32 million
Ricursive said total development cost was about $32 million. It said this covers pretraining trials, data mixing, evaluation, synthetic data generation, post-training development and the final training run. The company estimates this at about one-tenth that of previous top-level models, but did not give the amounts it is comparing against. A footnote on the announcement page noted reports that Reflection spends $150 million a month on computing resources.
The company also gave figures on the amount and efficiency of training. Pretraining data totaled 12 trillion tokens, less than Beam's 23.8 trillion tokens and Inkling's 45 trillion tokens. Reinforcement learning involved about 1.4 million trials, about one-seventieth of Beam's, it said. It did not buy outside data and used synthetic data it generated itself. It reported that in experiments at the 4-billion-parameter scale, computing efficiency was up to 16 times higher than the method it started from. All of these are the company's own statements and have not been verified by third parties.
API pricing of $0.20 for input; context window of 16,384 tokens
When ORIENS checked OpenRouter's listing on Nov. 5, the only provider was GMICloud, with pricing of $0.20 for input and $0.60 for output per 1 million tokens. Cached input costs $0.05. On the same OpenRouter, prices from the cheapest provider are $0.25 for input and $1.00 for output for Beam, $0.40 and $1.76 for Nemotron 3 Ultra, $1.00 and $4.05 for Inkling, and $1.25 and $4.25 for Muse Spark Open, and Ripple-65B Preview is cheaper than all of them.
On the other hand, this provider's context window is only 16,384 tokens, with output of up to 14,745 tokens. That is far below the maximum of 1 million tokens claimed in the model card, and also differs from the roughly 1.05 million tokens listed by Beam's provider. OrcaRouter noted that using long contexts currently requires running the weights in one's own environment.
- Ripple-65B Preview
- Beam
- Muse Spark Open (Meta)
Released under Apache 2.0, said to run on local devices
The license is Apache 2.0, allowing commercial use without a separate agreement. Muse Spark Open's own license covers users with revenue of up to $20 million and requires attribution on products.
The weights are about 130GB in BF16 and about 57GB in a compressed NVFP4 version. Goldie said it can run locally on a MacBook or Nvidia's DGX Spark. For data centers, the model card states it can generate 250 tokens per second on two "B200" GPUs.
According to the announcement page, it is offered on Together AI, GMI Cloud and OpenRouter in addition to Hugging Face, and Fireworks is also expected to join on release day. ORIENS could not confirm Together AI's pricing. The announcement page does not say when the full version will be released.
