Ripple-65B Preview

Ripple-65B Preview

Quick answer: Ripple-65B Preview is Ricursive Intelligence's first public model: a 65B-parameter dense Transformer pretrained on 12 trillion tokens with a 1,048,576-token (1M) context window, released November 6, 2026. It scores 80.52% on TerminalBench 2.1 (self-reported), which the lab reports as US open-weight state-of-the-art on agentic coding — ahead of significantly larger open-weight rivals including Reflection Beam 501B (80.1%), Meta Muse Spark Open 816B (76.4%), Thinking Machines Inkling 975B (63.8%), and NVIDIA Nemotron 3 Ultra 550B (56.4%). It is open-weight under Apache 2.0.

At a Glance

Where Ripple-65B Preview leads

• 80.52% TerminalBench 2.1 (self-reported) — ahead of every compared open-weight model regardless of size, per the lab's own comparison table
• Dense 65B architecture fits on a single NVIDIA DGX Spark, unlike the 500B+ MoE rivals it outperforms on agentic coding
• 97.1–98.3% across AIME 2024/2025/2026 — competitive math scores at a fraction of rivals' parameter count
• Reported $32M total R&D cost — about one-tenth the compute spend the lab attributes to the prior open-weight SOTA

Where it lags

• 89.6% MMLU and 84.2% MMLU-Pro trail all four larger MoE comparison models (86.8–87.9% MMLU-Pro)
• 85.2% GPQA Diamond is the lowest of the five compared models (others: 87.0–90.6%)
• FORTRESS Benign (over-refusal) at 95.5% is mid-pack; FORTRESS Adversarial (83.8%) trails Muse Spark Open (94.2%) and Reflection Beam (92.1%)

Best for: teams wanting strong agentic-coding and terminal-automation performance from a model small enough to self-host on a single workstation-class GPU, rather than a 500B+ parameter MoE.

What Ripple-65B Preview Is

Ripple-65B Preview is the first public model release from Ricursive Intelligence, a lab building AI-driven chip design tools. It is a dense (non-MoE) Transformer pretrained from scratch on 12 trillion tokens, using a "thin-and-deep" aspect ratio (narrow d_model of 5,120 across 80 layers) derived from an Olmo 3 baseline with architectural changes — gated attention, global NoPE, FFN activation clipping, and value residuals — aimed at improving training stability and reasoning at this depth.

The lab explicitly chose a dense architecture over MoE to reduce training/serving infrastructure complexity, reporting 40% MFU throughout pretraining and posttraining without custom kernels or quantization. Post-training focused narrowly on agentic coding (midtraining, long-context extension to 2M tokens internally, SFT, and RL), which the benchmark table reflects: Ripple-65B's agentic/coding scores are strongly competitive with or ahead of MoE rivals carrying 8–15x more total parameters, while its general-knowledge scores (MMLU, GPQA) trail those same larger models.

Ripple-65B Preview ships alongside a base-model checkpoint (Ripple-65B Base Preview, not separately tracked on this page) and is available in BF16 and NVFP4 numerics, plus third-party hosting via Together AI, Fireworks, GMI Cloud, and OpenRouter.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing from Ricursive Intelligence directly. Available on Hugging Face in BF16 and NVFP4, and hosted by third parties including Together AI, Fireworks, GMI Cloud, and OpenRouter (check those providers for current per-token rates).

Context Window

Ripple-65B Preview has a 1,048,576-token (1M) context window — roughly 1,500+ pages of text in a single request. The lab reports an internal long-context training stage extending to 2M-token sequences during posttraining, though the publicly stated max context is 1M.

Public Benchmark Scores

Scores above are reported by Ricursive Intelligence and shown for context. They are not Benchgen measurements. The lab notes Terminal-Bench 2.1 scores shift notably by harness/sequence length — it also reports an 80.5 Terminus-2-harness score (not tabulated here as a separate row; see model card) alongside the 80.52 self-reported figure.

Not mapped to an existing Benchgen benchmark (bucket C, documented per skill policy rather than guessed):

• SWE-bench Pro v2.0 Hard — a harder, dated subset that doesn't match our existing swe-bench-pro entity's general methodology; excluded rather than silently widened.
• Base-model pretraining evals (MMLU, BBH, HellaSwag, etc. for Ripple-65B Base Preview) — these describe a distinct pretrain-only checkpoint, not the instruct model this page covers; deferred to a possible future base-model page rather than conflated here.

Ripple-65B Preview vs Alternatives

Ripple-65B Preview trades roughly a 1–5 point deficit on pure knowledge recall (GPQA, MMLU) for leading agentic-coding throughput at 8–15x fewer total parameters than its closest rivals — a meaningfully cheaper self-hosting footprint for teams whose workload is terminal/coding agents rather than broad knowledge Q&A.

How Ripple-65B Preview Performs on Real Agent Tasks

Ripple-65B Preview's benchmark profile is narrowly optimized: strong on TerminalBench, SWE-bench Verified, MCP-Atlas, and τ³-bench banking — all agentic, tool-use, or terminal-automation evaluations — while trailing on knowledge-heavy tasks. This matches the lab's stated posttraining focus (a relatively light posttraining pass concentrated on agentic coding rather than broad capability). A leaderboard score establishes peak capability under the vendor's own harness and effort settings; it does not establish reliability across a team's specific tools, error-recovery behavior, or cost-at-scale, all of which matter more for production agent deployments than a single aggregate benchmark number.

Use Ripple-65B Preview via API

from openai import OpenAI  # swap SDK/base_url per provider, e.g. Together AI, Fireworks, OpenRouter

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.together.xyz/v1")

response = client.chat.completions.create(
    model="ricursiveai/ripple-65b-preview",
    messages=[{"role": "user", "content": "Write a shell script to rotate log files older than 7 days."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: Beam 1.0 501B-A23B · Nemotron 3 Ultra 550B-A55B
• Benchmarks: TerminalBench · SWE-bench Verified · GPQA Diamond

Frequently Asked Questions

Specs and scores sourced from Ricursive Intelligence's official Hugging Face model card and launch blog; third-party benchmark scores attributed inline. Last updated 2026-11-07.