Ripple-65B Preview

Research checked 6 November 2026. Publisher claims and third-party trials are identified separately; this page is not an AIDreamHub hands-on benchmark.

An open coding model, not a complete repository agent

Ripple-65B Preview is a 65B dense Transformer from Ricursive Intelligence, with post-training focused on agentic coding. The release includes instruction-tuned BF16 and NVFP4 variants and a separate base model. The linked BF16 repository is not interchangeable with the quantised release or a hosted endpoint.

It can be considered as the model behind a repository or operations agent. The surrounding system must still supply tools, repository state, permissions, execution and tests. A model score alone does not give it a safe shell or prove a pull request is correct.

Official project page captured by AIDreamHub; not a locally run product.

Read the launch benchmark with its harness

The publisher reports about 80.5 on Terminal-Bench 2.1 and 76.5 on SWE-bench Verified. It also explains that comparison values mix official reports and in-house measurements and that harness choices affect results. That makes the table useful evidence to investigate, not a universal ranking across all agents.

The current card describes up to 1Mi context, or 1,048,576 tokens. Supported context length does not mean that length is economical or equally effective on every runtime. Test retrieval and truncation behaviour at realistic lengths instead of sending a whole repository by default.

How to use: choose hosted qualification or a hardware-specific local path

For a small pilot, a hosted endpoint can separate model evaluation from local serving work. GMI documents an OpenAI-compatible chat endpoint and model identifier; the example below uses an environment variable for its key. Do not send private code without reviewing the provider’s data handling and your organisation’s permissions.

For local use, choose the exact BF16 or NVFP4 package and follow the release’s compatible serving instructions. Inspect custom code before trusting it. Pin revisions, begin with a short context and capture tool traces, tests and resource usage. We did not run local inference or reproduce the launch scores.

curl https://api.gmi-serving.com/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $GMI_API_KEY" --data '{"model":"gmicloud/Ripple-65B-Preview","messages":[{"role":"user","content":"Explain the failing test before proposing a patch."}],"max_tokens":800}'

A provider field report is useful—but not independent certification

GMI’s launch-week report describes small bug-fix repositories and a simulated incident workflow with traps and a human approval gate. It includes a single 100-agent run and repeated smaller runs. These are the serving provider’s measurements on its own endpoint, not proof of equivalent local performance or arbitrary production-code success.

The transferable lesson is its grading method: external tests determined whether a fix worked, while the model’s own summary had no grading weight. Approval was enforced by the tools. A few prompt-injection trials do not establish general immunity, so keep permissions and validation outside the model.

Budget memory before downloading

As an illustrative lower-bound calculation, 65 billion parameters at two bytes each are about 130 GB of weight data; ideal four-bit storage is about 32.5 GB. These are decimal arithmetic estimates, not measured deployment requirements. Quantisation metadata, activations, runtime buffers and context state add overhead.

Do not convert a DGX Spark example into a promise for any laptop, or a datacentre throughput number into single-user latency. Record prompt size, concurrency, quantisation, time to first token and accepted-task cost together. Longer context can change the dominant bottleneck.

An agent evaluation that can catch false success

Start with a fixed failing test in a disposable repository. Require the change, test output and final diff, then include a no-change task, an invalid tool response and a denied operation. Keep summaries separate from observed state. For operations, require a matching human approval before any production mutation.

Compare Ripple with your current model on the same tools, task set and budget, and compare deterministic scripts for work that does not need reasoning. Count retries and rejected patches. An endpoint that generates code quickly can still be expensive if people must repair its outputs.

License and deployment decision

The release is Apache-2.0-labelled, while hosting, integrations and source-code rights remain separate. Hosted pricing and limits can change; verify the selected endpoint rather than baking launch-week rates into a long-term cost promise.

The best first decision is whether this model improves accepted work in your existing execution loop. Only then decide whether local serving economics justify the hardware and maintenance. Our hand drawing explains capacity planning; it is not a performance chart or model result.

Questions before adoption

Does 65B fit a 32 GB GPU at four bits?

The ideal raw-weight arithmetic is about 32.5 GB in decimal units before metadata, buffers and context. It is not a fit guarantee. Use the release’s hardware-specific instructions and measure the actual quantised runtime with headroom.

Do launch scores prove it is the best coding agent?

No. A model and an agent harness are different components, and the publisher notes comparison and reproducibility limits. Evaluate your own repositories, tool contracts and acceptance tests at a fixed budget instead of extending a single leaderboard to all work.

Can a successful tool call replace approval and tests?

No. The executor should enforce permission, validate arguments and verify the resulting artifact or system state. Provider demos that respected approval used an actual tool-side gate; model compliance alone is not the safety guarantee.

Sources reviewed

• Ricursive Intelligence — model card
• Ricursive Intelligence — launch and evaluation conditions
• GMI Cloud — provider field tests

Explore related tools

• Coding tools
• Text models