How Much To Run AI
Open weights

Ripple-65B Preview: API Costs and Local Deployment

Ripple-65B Preview is Ricursive Intelligence’s dense, open-weight text model for reasoning and coding agents, with support for up to a 1,048,576-token context. Start with a hosted endpoint for an occasional task, or evaluate the instruction-tuned weights on suitable hardware. Its 65B size makes quantization important: the experimental NVFP4 release quantizes the MLP layers while keeping attention in BF16.

Content reviewed:

Use this model

API docs and console

Official model repository

Use cases

  • Try a coding agent on a reproducible bug with tests that verify the patch.
  • Compare hosted token billing with the cost of keeping a local server available.
  • Choose instruction-tuned BF16 or NVFP4 for an application; use Base for continued pretraining or fine-tuning.

How to use Ripple-65B Preview

Ricursive links hosted access through Together AI, GMI Cloud and OpenRouter. For the OpenRouter route, use ricursiveai/ripple-65b-preview with an account API key. The example below sends a text request; a coding agent additionally needs a client that executes tools and returns their results.

Begin with a short explanation task, then connect a coding client to a disposable test repository. Set a token and iteration budget, inspect the patch and run the tests before accepting completion. Include unsuccessful attempts when comparing models.

Illustrative OpenRouter request; sending it incurs usage charges
import os
from openai import OpenAI

client = OpenAI(base_url="https://openrouter.ai/api/v1",
                api_key=os.environ["OPENROUTER_API_KEY"])
response = client.chat.completions.create(
    model="ricursiveai/ripple-65b-preview",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Propose tests for a CSV importer that rejects duplicate IDs."}],
)
print(response.choices[0].message.content)

Ripple API pricing and a task-budget example

OpenRouter’s GMI Cloud listing checked November 5, 2026 quotes $0.20 per million input tokens, $0.60 per million output tokens and $0.05 per million cache-read tokens. Record the selected provider and actual usage alongside the bill; retries and every model turn contribute to a coding task.

At those fresh-input and output rates, 100,000 input tokens plus 10,000 output tokens cost $0.026 before external tools and other charges: $0.020 input plus $0.006 output. This is a token-budget example, not a measured agent run. A one-million-token context is capacity, not an included allowance.

Choose Instruct BF16, NVFP4 or Base

The release includes instruction-tuned BF16, experimental NVFP4 and a separate BF16 Base model. NVFP4 uses a mixed format: the MLP is quantized and attention remains BF16. A generic four-bit-per-parameter calculation will therefore understate this artifact’s weight storage.

As a weight-only arithmetic reference, 65 billion parameters at two bytes each take about 130GB decimal before cache, activations and runtime buffers. The publisher identifies DGX Spark as a local target for the quantized model. Select the matching serving recipe and measure available context and peak memory rather than treating the advertised 1M limit as a device guarantee.

DFlash2 is a separate draft model for speculative decoding. It adds memory and requires a compatible execution path. Evaluate the combined setup on code and prose separately because acceptance rates can change generation speed.

What does the TerminalBench score establish?

Ricursive reports 80.52 on TerminalBench 2.1 and describes it as a US open-weight agentic-coding lead. Its comparison scope is US open-weight models, rather than every coding model. The card describes BF16 evaluation and the harness choices; a score from another TerminalBench version is a different measurement.

Use that result to select a trial workload. For the deployment decision, compare accepted patches, total tokens and elapsed task time under the same client and test suite. The publisher’s $32 million figure covers development and training, not inference pricing or the cost of your deployment.

When does a local server make sense?

A local deployment is useful when you need control over data, model files and serving. For occasional tasks, first compare the low hosted token bill with GPU reservation, storage and setup time. For sustained traffic, measure completed work per reserved hour at the intended concurrency.

For example, a server assumed to cost $2 per hour and reserved for 30 minutes costs $1 before storage and operations. That rental assumption does not establish Ripple throughput. Divide the complete bill by accepted results, including failed attempts and idle reservation.

Ripple weights use Apache 2.0

The official BF16 and NVFP4 repositories label the release Apache 2.0. Retain the supplied license and notices when redistributing it. Hosted services have their own account, availability and billing terms.

Before you use it

  • A model endpoint generates tool requests; your agent client controls execution and verifies results.

How to get started

Start with the hosted endpoint or a matching instruction-tuned checkpoint, then compare the cost and time per patch that passes your tests.

Sources

Sources (4)

Calculators