TL;DR
Ricursive’s Ripple-65B Preview challenges larger open-weight coding models on a vendor-reported terminal benchmark. Its real lesson is about evaluation trade-offs, not a universal win for smaller models.
A new open-weight contender
Ricursive Intelligence has released Ripple-65B Preview, an open-weight model designed for agentic coding. Its benchmark claims invite a useful question: can a smaller, dense model compete with much larger mixture-of-experts systems when an agent has to use a terminal? The release arrived on November 4, according to coverage of the launch. (Source : MarkTechPost — Ripple-65B Preview)
Ricursive says Ripple is a 65-billion-parameter dense Transformer pretrained on 12 trillion tokens, with support for up to one million tokens of context. The instruction-tuned checkpoint is available on Hugging Face under Apache 2.0; the release also includes an experimental quantized variant and a separate base model. (Source : Ricursive — Ripple model card; Ricursive — NVFP4 checkpoint)
The benchmark lead has boundaries
Ricursive reports 80.52 on TerminalBench 2.1 for Ripple, compared with 80.1 for Reflection Beam in its comparison. That is a narrow 0.42-point difference, not proof of a substantial lead or of superiority across all coding tasks. (Source : Ricursive — Launch announcement)
The comparison covers US open-weight models, not every model worldwide. More importantly, the model card combines published scores with the company’s own evaluations where official results are missing. Ricursive says benchmark replication is difficult and that harness and context choices can change results. Independent runs with matching environments, repeated trials and uncertainty estimates would make a close result more informative. (Source : Ricursive — Evaluation methodology)
The vendor’s own table also shows why a single leaderboard is insufficient. Ripple scores 76.5 on SWE-bench Verified while Beam scores 81.6 in that table. On DeepSWE 1.1, Ripple’s 44.3 trails Muse Spark Open’s 60.0 in Ricursive’s comparison. These scores do not establish how either model will perform on a particular team’s codebase; they show that the answer depends on the task. (Source : Ricursive — Instruct model benchmarks)
For a different perspective on agent evaluation, see our analysis of Microsoft’s Agensh experiments: the setup and measured outcome matter as much as the headline.
Dense versus sparse: compare active work, not just size
Ripple activates its 65 billion parameters as a dense model. Reflection describes Beam as a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active. Comparing only 65B with 501B can therefore mislead: total weights affect storage, but active parameters, memory bandwidth, batching and implementation influence the work needed to generate tokens. (Source : Reflection — Introducing Beam; Ricursive — Model card)
Ricursive reports 250 tokens per second on two NVIDIA B200 GPUs at batch size eight, before quantization or custom kernels. That is a vendor measurement for a particular configuration, not a latency guarantee for a single local user. (Source : Ricursive — Model card)
What the $32 million figure does and does not say
Ricursive says its $32 million figure covers full R&D from scratch, including experiments, evaluation, synthetic data and final training. The company estimates this at one-tenth the cost of the previous open-weight agentic-coding leader. Its own explanation bases the rival estimate on reported compute commitments and training duration, so the ratio is an estimate rather than an audited, like-for-like cost comparison. (Source : Ricursive — Cost and methodology)
The distinction matters for buyers: training expenditure does not determine their per-task serving cost. A meaningful production comparison would measure dollars per accepted change, tokens consumed, retries, completion time and human review under the same tasks and tooling.
What developers can test now
The model card provides BF16 weights and links an experimental NVFP4 checkpoint. Ricursive positions the quantized model for hardware such as NVIDIA DGX Spark, but fit and throughput depend on the runtime, available memory and context length. The NVFP4 model card says it quantizes MLP layers while keeping attention in BF16. Developers should not assume it reproduces the full-precision benchmark exactly. (Source : Ricursive — Model card; Ricursive — NVFP4 checkpoint)
Start with an isolated coding environment, a fixed task suite and explicit review of generated changes. Compare the same harness and resource budgets against an existing model before treating a benchmark advantage as a deployment decision.
FAQ
Are Ripple’s weights available?
Yes. The instruction-tuned BF16 checkpoint is on Hugging Face under Apache 2.0, alongside links to the base and experimental quantized variants.
Is Ripple the best coding model overall?
The reported terminal score is competitive within Ricursive’s stated US open-weight comparison. Other benchmarks in its table favor different models; independent replication is still important.
Does 65B mean cheaper inference than a 501B model?
Not necessarily. Ripple is dense while Beam is sparse, so active parameters and the actual serving setup matter alongside total weights.
Does a one-million-token context guarantee good long-context performance?
No. A supported context length is a capacity limit, not a guarantee of accuracy or affordable latency at that length.
Further Reading
- Ricursive: Ripple-65B launch and cost discussion
- Ricursive: model card and benchmark methodology
- Reflection: Beam architecture and training
- Ricursive: experimental NVFP4 checkpoint
— The Agent Report