Ricursive Intelligence dropped Ripple-65B Preview on November 4 — a 65-billion-parameter dense model that reportedly tops open-weight models 7 to 15 times its size on agentic coding benchmarks. The compute bill: $32 million, roughly one-tenth what the previous open-weight leader cost to build. Apache 2.0 weights are live on Hugging Face today. The team behind it built AlphaChip, the reinforcement learning system that reduced Google TPU floor planning from months to hours.
The Benchmark Numbers and the Honest Caveats
Ricursive’s own evaluations put Ripple-65B at 80.5% on Terminal-Bench 2.1, 79.1% on SWE-Bench Pro v2.0-Hard, and 44.3% on DeepSWE v1.1. On Terminal-Bench, that is a 0.4-point edge over Reflection Beam-501B (80.1%) and a 4-point lead over Meta Muse Spark Open (816B, 76.4%). On paper, a 65B model beating an 816B model is a headline worth reading.
| Model | Parameters | Terminal-Bench 2.1 |
|---|---|---|
| Ripple-65B | 65B | 80.5% |
| Reflection Beam-501B | 501B | 80.1% (Ricursive) / 70.8% (official harness) |
| Meta Muse Spark Open | 816B | 76.4% (Ricursive) / 78.7% (official harness) |
| Thinking Machines Inkling-975B | 975B | 63.8% |
| NVIDIA Nemotron 3 Ultra 550B | 550B | 56.4% |
When Ricursive re-ran rivals on the official Terminus 2 harness at 128K context, Beam dropped to 70.8% and Muse Spark Open fell to 78.7%. Ripple held at 80.5%. The lead widens under that configuration, but re-running competitors raises fair questions about methodology. The community noticed: a heavily-upvoted Hacker News comment called the “US open-weight SOTA” framing “category narrowing to this degree” that has “diluted the term quite a lot.” The absent comparison is Qwen, which remains the global coding benchmark leader but carries Chinese-origin restrictions for some enterprises.
The benchmark lead is real. It is also narrow, self-measured, and scoped to US-origin models. Keep both things in mind.
Dense Architecture, by Design
Most frontier models use Mixture-of-Experts, routing each token to only a subset of parameters to keep inference costs down. Ricursive went the other way. All 65 billion parameters are active on every token.
The architecture is thin-and-deep: 80 layers, d_model 5120, inspired by Gemma 3N. Training ran in plain BF16, no custom kernels, no quantization, achieving around 40% model FLOPs utilization. That is a respectable figure for a five-person team. The choice of dense architecture from an AlphaChip-pedigree team is deliberate. When your founders optimize chips for a living, hardware utilization is not an afterthought.
Running It Locally: What You Actually Need
The marketing says “runs on a MacBook.” Technically true, but only on a MacBook Pro with 128 GB unified memory or higher. The Q4_K_M GGUF quantization weighs around 40 GB, and once you account for macOS overhead and the KV cache, a 48 GB MacBook Pro is too tight. You need 64 GB minimum; 192 GB (M3 Ultra Mac Studio) is comfortable. The two fastest paths:
# Ollama (v0.40.16+)
ollama run ripple-preview:65b-q4_K_M
# llama.cpp (b12167+)
llama serve -hf AtomicChat/Ripple-65B-Preview-GGUF:Q4_K_M
On DGX Spark, the NVFP4 checkpoint (~57 GB) runs with a provided script, but capped at 128K context. The advertised 1 million token context window requires two B200 GPUs in BF16. “1M context” is a specification, not a DGX Spark feature out of the box.
Who This Is Actually For
The clearest use case is organizations that need a capable open-weight coding model but cannot use Qwen: government contractors, defense-adjacent teams, and companies operating under US export control guidance. Apache 2.0 means unrestricted commercial use, modification, and redistribution. For teams with that constraint, Ripple is currently the strongest available option.
Beyond compliance, there is a real case for teams running an M3 Ultra or DGX Spark who want a production-quality agentic coding model that keeps data on-device. The model was validated against Ricursive’s internal chip-design platform before release, a demanding workload by any measure. The base model weights are available, and community fine-tunes are already under way on Hugging Face.
If you are on a standard developer MacBook, waiting for smaller distillations is the smarter move. If you are running 192 GB of unified memory and want a reason to use it, the Ollama page has everything you need. For a thorough breakdown of the benchmark methodology, this analysis from ibl.ai is worth reading before drawing conclusions.
The Takeaway
Ripple-65B is a legitimately interesting result: a five-person team with chip-design expertise building a 65B dense model that challenges systems 7 to 15 times its size at around one-tenth the reported compute cost. The benchmark story has caveats worth knowing. The hardware requirements are steeper than the marketing implies. For teams without a Qwen restriction, it is not automatically the best choice.
For what it is — an Apache 2.0, US-origin, locally-runnable agentic coding model with credible lineage from the AlphaChip team — Ripple-65B is the most interesting open-weight coding release of November. Ricursive says this is a Preview; posttraining is ongoing. The weights are on Hugging Face now.

