Ripple-65B Preview: 65B Dense Open Model, Benchmarks, Apache 2.0

November 6, 2026 | AI | By Mahesh

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, an Apache 2.0 open-weight language model with 65 billion dense parameters trained from scratch on 12 trillion tokens.[1] The chip design startup says the model scores 80.52 on Terminal-Bench 2.1, the highest figure among US open-weight models, while carrying a fraction of the parameters of the rivals it passes.[1][3] The weights are on Hugging Face, and Together AI serves the model through an API.[1][2]

What is new in Ripple-65B Preview

Ripple-65B Preview is the first public language model from Ricursive Intelligence, a general-purpose variant of the model behind its AI chip design platform and tuned for agentic coding.[2] It is dense, so all 65 billion parameters run on every token, unlike the mixture-of-experts rivals it is compared against.[1] Ricursive says it chose a dense design to cut training and serving complexity.[1]

The release has three checkpoints: an instruction-tuned BF16 model, an experimental NVFP4 build aimed at local hardware such as the NVIDIA DGX Spark and a BF16 base model that is pretrain-only.[1] Ricursive says the Preview label reflects relatively light post-training with a focus on agentic coding, so the instruct model should be read as a first cut.[1]

"our 65B dense LLM we pretrained from scratch"

Ricursive Intelligence, Official Announcement on X [3]

Ricursive also makes efficiency claims. It reports a 16x compute efficiency gain on code over its Olmo 3 baseline ladder, rising to 55x for the full run. It also cites a $32 million compute budget, which it calls one tenth of the cost of the previous open-weight leader in agentic coding.[1][3] Those cost figures come from the company and have no independent audit. The team grew from one to five people during development.[1]

Under the hood: architecture and specifications

Ripple-65B Preview is a 65 billion parameter dense Transformer with 80 layers, a model dimension of 5,120 and support for up to 1,048,576 tokens of context.[1][2] The table below collects the published specifications.

The design departs from the Olmo 3 llama-like baseline. Ricursive moved to a Gemma 3N-like architecture for training stability at scale, added more attention heads than usual, raised the ratio of sliding-window to global attention layers and shrank the window to offset the cost.[1] It also added FFN activation clipping, gated attention, global NoPE and value residuals.[1]

Pretraining ran for 877,500 steps at 8K sequence length, followed by a 200 billion token midtraining stage at 64K, long-context training, 183 billion tokens of supervised fine-tuning and 700 steps of reinforcement learning at 128K.[1] Together AI reports 1.4 million reinforcement learning rollouts behind the Terminal-Bench result.[2] Ricursive says it held 40% model FLOPs utilization with simple sharding, no custom kernels and no quantization, and reports 250 tokens per second on two B200 GPUs at batch size 8 before any optimization.[1]

Benchmarks and performance: reported results against marketing claims

Ricursive reports 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4.[1] The comparison covers US open-weight models, and Ricursive ran all of them at full BF16 precision and maximum effort.[1]

Source: Ricursive Intelligence model card. Some rival scores come from Ricursive's own evaluation runs rather than official model cards.[1]

The card carries a caveat that matters for anyone reproducing the headline number. Ricursive found that Terminal-Bench 2.1 scores came out lower for nearly every model, Muse Spark Open excepted, when using the official Terminus 2 harness at 128K sequence length with summarization.[1] It describes its table as a best-effort fair comparison, and it filled missing rival scores with its in-house evaluation flow.[1]

Outside terminal work the picture is mixed. Ripple leads SWE-bench Pro v2.0 Hard at 79.1 against 76.5 for Reflection Beam, yet scores 44.3 on DeepSWE 1.1 against 60.0 for Muse Spark Open and 39.4 on SciCode against 58.5.[1] On competition math it posts 98.3 on AIME 2025 and 95.0 on AIME 2026.[1] No independent evaluation had surfaced at the time of writing, so every figure here is vendor-reported.

Pricing, availability and how to access Ripple-65B Preview

Together AI lists Ripple-65B Preview at $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.02 per million.[2] The endpoint name is ricursive-ai/Ripple-65B-Preview, and Together supports reasoning and tool calling for agentic coding workflows.[2] Its spec sheet shows a 128K context window even though the model supports up to 1M, so check each provider's limit before planning long-context work.[2]

Developers who want to self-host can download the BF16 instruct weights, the NVFP4 build or the base model from Hugging Face under Apache 2.0.[1][4] A BF16 copy of 65 billion parameters is about 130 GB by simple arithmetic, which fits on a multi-GPU server and exceeds a single consumer card. Ricursive points to the NVFP4 build and a Colab notebook for local experiments.[1] Ricursive's launch post also names OpenRouter and GMI Cloud as API partners.[3]

Limitations, trade-offs and competitive alternatives

Ripple-65B Preview is a text-only, mostly English model with light post-training, and its Hugging Face repository carries a custom code tag, which means loading it may require trusting remote code.[1] The base model also trails DeepSeek V4 Flash and Nemotron 3 Ultra on several traditional pretraining evaluations in Ricursive's own base table, such as MMLU-Pro and MATH.[1] Teams that need strong tool use should note the MCP Atlas gap of 71.67 against 89.2 for Muse Spark Open.[1]

Larger open-weight options include Tencent's 770 billion parameter model in Tencent Hy4: 770B MoE, 1M Context, 6 Yuan per Million Tokens and the 1.05 trillion parameter preview in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing. Moonshot's 2.8 trillion parameter Kimi K3.1 Pro also shipped open weights this week. Ripple is the option to test when serving cost and a permissive license matter more than the last few points on every benchmark.

Sources

1. Ricursive Intelligence. "Ripple-65B Preview model card, benchmarks and training details," Hugging Face, November 4, 2026. huggingface.co and launch blog
2. Together AI. "Ripple-65B Preview API: Pricing, Benchmarks and Docs." together.ai
3. Ricursive Intelligence. Launch announcement thread on X, November 4 and 5, 2026. x.com
4. Ricursive Intelligence. "Apache 2.0 license and NVFP4 checkpoint," Hugging Face. license and NVFP4 repository

Read More on Depth Grid

• Tencent Hy4: 770B MoE, 1M Context, 6 Yuan per Million Tokens
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Grok Voice 3.0 Launch: $0.08 per Minute, 86.9 Index, Specs
• AI Native Startup Playbook: Six Moves From Idea to Series A

Article by Mahesh | Depth Grid