What happened
Ollama’s library page describes Ripple-65B Preview as a 65-billion-parameter dense model from Ricursive Intelligence, built for agentic coding. It says the model was pretrained from scratch on 12 trillion tokens, with all 65 billion parameters active on every token. The page characterizes it as a general-purpose version of the model behind Ricursive’s AI chip-design platform.
The listed workflow starts with ollama run ripple-preview. Ollama also documents an HTTP chat API example and Python and JavaScript examples using its client libraries. The page lists launch integrations for Claude Code, OpenCode, Hermes Agent and OpenClaw, each with an ollama launch command that specifies ripple-preview as the model. That makes the release more than a model card: Ollama presents paths to place it into existing coding-agent workflows.
Ollama says Ripple was trained with large-scale reinforcement learning to carry out multi-step terminal tasks like a coding agent. It advertises a context window of up to 1 million tokens. The model listing shows ripple-preview:latest and ripple-preview:65b at 39–46 GB, with a 2 million-token context figure in the listing table; that differs from the “up to 1M” description elsewhere on the same page. Builders should treat the precise supported context as something to verify for the selected build and runtime rather than infer from a single headline number.
The page says Ripple uses DFlash 2 speculative decoding on Ollama’s MLX engine to generate faster, without giving a speed figure. It identifies the license as Apache 2.0 and says commercial use is allowed. On benchmarking, Ollama attributes a 80.5 score on Terminal-Bench 2.1 to Ricursive and links to the company’s launch post for details.
The catalog page shows 1,324 downloads and says it was updated 21 hours earlier at the time shown. Those figures establish that this is an active library listing, not an announcement of a separate model version. They do not establish production adoption, nor do they reveal how many downloads led to a successful local coding-agent deployment. The page also labels the listed variants as text models, while its examples show chat-style calls through Ollama’s API and client libraries.
What the archive adds
New Runtime covered Ricursive’s Ripple-65B Preview release on 4 November 2026, including its reported TerminalBench result and development-cost figure. This Ollama listing is a new access and integration development on that known model story, not a new model announcement. An earlier note also covered GPTQModel’s Ripple quantization support, but the central change here is Ollama’s direct launch path.
Trend signal
This supports agent economics moves to completed work only weakly: the model is explicitly trained for multi-step coding tasks, but Ollama’s page supplies no cost-per-verified-task or comparative completion evidence. It is an access signal, not evidence that the economics of completed work have shifted.
Why now
The Ollama page lists the model as updated 21 hours before the page was captured, marking a recent addition to its library. The source does not explain why the listing happened at this point or describe a broader rollout schedule. Its placement alongside launch commands for established agent tools means the immediate change is operational availability rather than another stated training or benchmark milestone.
Connections to other events
New Runtime’s earlier coverage of Ripple’s original release described the same 65B dense model as an open-weight option for agentic coding. The later GPTQModel note concerned quantization support and checkpoint resharding. Ollama now adds a separate distribution and runtime route, so these are complementary steps: model release, tooling support, and a directly documented local launch path.
Why it matters
A local launch route can make an open-weight coding model easier to test inside an existing agent workflow, and Apache 2.0 removes a licensing barrier for commercial experimentation. Teams can evaluate it against a familiar toolchain rather than building a bespoke serving path first. The documented Claude Code, OpenCode, Hermes Agent and OpenClaw commands are particularly useful as starting points for controlled trials.
The practical trade-off is substantial: the listed 39–46 GB size can constrain who can run it locally, while the advertised long context and coding-agent training do not by themselves establish useful performance on a team’s repositories. A long-context claim may matter for large codebases or lengthy sessions, but quality, latency and memory pressure remain workload-specific. The reported Terminal-Bench score is attributed to Ricursive, so it should be treated as a vendor claim until teams inspect benchmark details and reproduce results relevant to their own tasks.
Limits and uncertainty
Ollama’s page gives conflicting context figures: up to 1 million tokens in the feature description and 2 million in the listing table. It does not specify which hardware configurations achieve the listed model size range, or provide measured latency, throughput, memory use, or comparative results for the DFlash 2 decoding claim. The Terminal-Bench 2.1 score is reported by Ricursive; independent reproduction and details of evaluation conditions are not established by the listing. Teams should test compatibility, quality and resource use in their own setup.
Skill to build now
Run a bounded evaluation before changing a coding-agent default. Install the listed model with ollama run ripple-preview, then use one representative repository and a fixed set of multi-step terminal tasks. Record whether the selected build loads, its practical context ceiling, elapsed task time, memory use, and the rate of accepted changes after review. Compare those results with the model already used by the team; do not treat a benchmark score as a substitute for this test.
Reader takeaway
Ripple-65B Preview is now easy to try locally through Ollama, including from several coding-agent tools, and its Apache 2.0 license allows commercial use. Treat it as a substantial hardware commitment and validate the context limit and coding performance on your own workloads before relying on it.
Open question
Which hardware and Ollama configurations support the advertised context limits at useful speed?