Ripple-65B tops open models five to fifteen times its size on agentic coding
On November 4, 2026, Ricursive released Ripple-65B Preview, a dense 65-billion-parameter open-weight model trained from scratch for agentic coding: 80.5 on Terminal-Bench 2.1, ahead of MoE rivals of 501 to 975 billion parameters, on $32 million of research spend. The scores remain self-reported and uneven outside coding; the interest is first and foremost an efficiency story.
November 4, 2026. Ricursive Intelligence releases Ripple-65B Preview, a dense 65-billion-parameter open-weight model trained from scratch for agentic coding. November 4, 2026. The announced score is 80.52 on Terminal-Bench 2.1, the highest self-reported among US open-weight models. November 5, 2026. The weights are on Hugging Face under Apache 2.0. The claim fits in one sentence: a dense 65-billion-parameter model beats 501- to 975-billion-parameter MoEs on agentic coding, at a tenth of the cost.
For a team that self-hosts its own coding agent, this is not just another release. It is the efficiency-versus-scale question posed in numbers, from an unexpected name — a lab that builds chips.
Who Ricursive is, and why its model came out of chip design
Ricursive Intelligence is not a consumer model vendor. It is a full-stack lab that co-optimizes hardware, software, and models for chip design. Ripple-65B Preview is the general-purpose variant of the proprietary model behind its AI-assisted chip design platform.
That origin explains two choices.
- A dense Transformer rather than an MoE. Ricursive rejected Mixture-of-Experts to cut infrastructure complexity, not to win a benchmark.
- A controlled budget. The announced total R&D cost is $32 million, roughly one-tenth of the prior open-weight agentic-coding leader, according to the lab.
The result is a thin and deep model: a model dimension of 5120 across 80 layers, trained on 12 trillion tokens, with context up to 1 million tokens. Training held 40% MFU with simple sharding, no custom kernels, and no quantization.
A recipe built around efficiency
The architecture choices tell the same story as the budget: simplicity as a lever.
- Overprovisioned attention heads for reasoning, offset by more sliding-window layers and smaller windows.
- Gated attention, global NoPE, and value residuals — three mechanisms borrowed from recent literature.
- FFN activation clipping, a technique also used in OpenAI’s gpt-oss.
The recipe was built through laddering experiments starting from Olmo 3’s ladder. Ricursive rejected any change that failed to hold at the 4-billion scale, such as two-stream multi-head causal. The technical post reports a 16× code compute-efficiency gain at that scale.
Post-training follows the same frugality. It combines truncated importance sampling, score centering, and a linear length penalty, with no purchased data — the synthetic pipelines were built in-house. Ripple reached its 80.5 with 1.4 million rollouts, roughly 70× fewer than the 100 million reported by Beam.
That frugality is the real signal in the announcement. If the score holds up, the lesson is not “small models beat big ones” but that training discipline counts as much as scale — and that a team can reach competitive agentic coding without lining up a rack of H100s.
The claim: 80.5 on Terminal-Bench, ahead of models five to fifteen times larger
The central benchmark is Terminal-Bench 2.1, which measures the ability to drive a real terminal to complete a task. Ripple-65B Preview reports 80.5, against:
- Beam from Reflection: 80.1, on 501 billion total parameters;
- Meta’s Muse Spark Open: 76.4, on 816 billion;
- Thinking Machines’ Inkling: 63.8, on 975 billion;
- NVIDIA’s Nemotron 3 Ultra: 56.4, on 550 billion.
Every rival it beats is an MoE with 501 to 975 billion total parameters. The ratio is brutal: 5 to 15 times more total parameters for a score that is lower or barely higher.
One detail shades the picture. When Ricursive re-ran Terminal-Bench 2.1 with the official Terminus 2 harness at 128K context, most rivals dropped — Beam fell to 70.8 — while Muse Spark Open rose to 78.7, still below Ripple’s 80.5.
The weaknesses: a coding specialist, not a generalist
The other side is just as clear. Ripple-65B Preview is an agentic-coding specialist, not a generalist model.
- SciCode: 39.4, the lowest of the five models in Ricursive’s own table.
- DeepSWE 1.1: 44.3, behind the pack.
- SWE-bench Verified: 76.5, below Beam (80.9) and Inkling (77.6).
- GPQA Diamond: 85.2, behind Muse Spark Open (89.0) and NVIDIA (87.0).
In mathematics, by contrast, the model holds up: 98.3 on AIME 2025 and 90.5 on HMMT February 2026. The profile is coherent — strong on task-oriented reasoning, weaker on pure scientific knowledge. It is the trade-off of a model trained to code, not to do everything.
One methodological limit also deserves mention: every score is self-reported, including those of the competitors cited. No independent evaluation has yet confirmed the 80.5 on Terminal-Bench. Prudence requires reading that number as a vendor claim, not a third-party verdict.
What this changes for the open-model landscape
The story Ripple-65B tells goes beyond the lab. It is about the cost curve of open-weight training.
A dense 65-billion model rivaling 500-billion MoEs — if confirmed — means a team can reach a competitive level without investing in MoE infrastructure. That is a signal for everyone hesitating to self-host a coding agent.
- Inference cost. The model reports 250 tokens per second on 2× B200 in BF16, and an NVFP4 variant targets the NVIDIA DGX Spark — a desktop machine.
- License. Apache 2.0, no commercial restrictions, with three checkpoints (instruction-tuned BF16, base BF16, NVFP4).
- Distribution. Weights on Hugging Face, inference on Together AI, GMI Cloud, and OpenRouter.
For a team that wants to keep its models in-house, the efficiency argument is concrete: a competitive coding agent that runs on one or two cards instead of a rack. That is the promise the MoEs made hard to keep.
The timing also matters. The open-weight coding tier has spent the past year racing toward larger MoEs — Beam, Inkling, Nemotron 3 Ultra — on the assumption that agentic coding rewards total parameter count. Ripple challenges that assumption at its root, at the exact moment when inference cost, not training cost, has become the binding constraint for teams that run models themselves. If a 65-billion dense model can do the job on one or two cards, the economic case for a 500-billion MoE gets harder to make.
Verdict
Ripple-65B Preview is an efficiency claim, not yet a confirmed revolution. Its headline score is self-reported, and its results remain uneven — excellent at agentic coding, weak at scientific knowledge.
If you self-host a coding agent and hardware cost is your constraint, Ripple-65B is worth a try: a dense 65-billion model approaching 500-billion MoEs, under Apache 2.0, with a desktop-class variant. If you need a generalist or your tasks touch scientific code, wait for independent evaluations. The lesson is already in: total parameter count is no longer the right proxy for competence, and agentic coding is becoming the first field where efficiency beats scale.