Ripple-65B Preview: a 65B-parameter agentic coding model under Apache 2.0 for $32M of compute

Ripple-65B Preview: a 65B-parameter agentic coding model under Apache 2.0 for $32M of compute

🔎 A dense, open model — and a bill that hurts the giants' egos

On November 4, 2026, the Palo Alto-based lab Ricursive Intelligence released Ripple-65B Preview: an agentic coding model with 65 billion parameters — dense, not MoE — pre-trained on 12 trillion tokens, with context of up to 1 million tokens. All of it, according to the lab, for roughly $32 million in total training cost. One-tenth the cost of the previous reference open-weight model for agentic coding, the company claims.

But the real news isn't the score on Terminal-Bench 2.1. It boils down to three words: Apache 2.0, on all checkpoints. Concretely, an agentic coding model that you can download, modify, fine-tune, embed in a commercial product and resell — without negotiating a license, without paying royalties, without depending on an API whose price can double overnight.

At a time when closed frontier models cost hundreds of millions of dollars, and open models tend to become giant MoEs of 500 to 800 billion parameters that are impossible to self-host, Ripple-65B makes a different bet: a dense architecture optimized for Nvidia hardware, a size that fits on reasonable hardware, and a license that leaves no room for the lawyers. Whether the numbers actually hold up remains to be seen — because, as we will see, they all come from the vendor and none have been independently replicated.

The Essentials

• Ripple-65B Preview is a 65B-parameter dense agentic coding model (no MoE), pre-trained on 12 trillion tokens, with context up to 1M tokens, released on November 4, 2026 by Ricursive Intelligence (source).
• Apache 2.0 license on all checkpoints: commercial use, modification, and redistribution allowed. This is the game-changer for enterprises.
• Claimed training cost: ~$32M for the entire process (experimentation, data pipeline, evaluations, post-training), not just the final run (source).
• 80.52 on Terminal-Bench 2.1 (self-reported), ahead of much larger US open-weight models like Reflection Beam 501B (80.1) or Muse Spark Open 816B (76.4). SWE-bench Verified: 76.5 (source).
• But: the model doesn't catch up with the leaders on the hardest benchmarks (SWE-Bench Pro v2.0 Hard: 79.1% vs 86.3% for Inkling 2 Small), and none of the numbers have been independently replicated (source).
• Hosted inference: $0.20 / $0.60 per million tokens (input/output) — very competitive (source).
• Deployable on a single DGX Spark in NVFP4 format, or on 2x B200 in BF16 with SGLang.

Recommended tools

Ripple-65B in numbers: what exactly does Ricursive claim?

Direct answer: a dense 65B-parameter model, 12 trillion pre-training tokens, 1M context, 80.52 on Terminal-Bench 2.1, 76.5 on SWE-bench Verified — and all of that for $32M.

The official Hugging Face card (ricursiveAI/Ripple-65B-Preview) details the scores:

The table deserves a careful read. On Terminal-Bench 2.1, the advantage over Reflection Beam is only 0.4 points — a gap that, on an agentic benchmark, can fall within statistical noise. And as the critical analysis from eesel AI points out, Ricursive only leads on 2 of the 22 benchmarks in its own table. The model is excellent at terminal coding, decent on SWE-bench, and clearly behind the leaders on the hardest tasks.

Another important detail: the post-training is "relatively light, focused on agentic coding," with 1.4 million rollouts — roughly 70 times fewer than Reflection Beam — and zero purchased post-training data. This is an unapologetically efficiency-driven recipe, not an attempt to dominate everything.

Why the Apache 2.0 license is the real news

Direct answer: because it turns a good coding model into a marketable asset. An open-weight model with a restrictive license remains a research artifact; an Apache 2.0 model is a product component.

Many recent open-weight models come with conditional licenses: commercial use restrictions beyond a user threshold, naming obligations, or non-compete clauses. Apache 2.0 leaves none of that: you can fine-tune the model on your proprietary codebase, integrate it into a SaaS, sell it as a service, and owe nothing to anyone.

For a team building a coding agent, this changes three things:

1. Marginal cost becomes predictable. You host the model, you pay for the GPU, period. No unilateral price changes, no rate limits imposed by a provider.
2. Compliance becomes tractable. Your data never leaves your infrastructure. For regulated industries, this is often the deal-breaker.
3. Modification is legal and encouraged. You can continue post-training on your own rollouts, exactly as Ricursive did with its 1.4M rollouts.

It's the same logic driving the open-weight ecosystem to accelerate on all fronts: Nvidia Nemotron 3.5 Super Lightning explores open hybrid Mamba-Transformer architectures, and Tencent opened Hy4 with 770B vision-native parameters and computer-use capabilities. Ripple-65B stands out in this landscape with a radical choice: staying dense and compact rather than joining the race for giant MoEs.

$32M: what a modest compute budget can (and cannot) achieve

Direct answer: $32M is enough for a state-of-the-art model in a narrow domain (agentic coding), but not for a frontier generalist model. The budget buys specialization, not universality.

First, let's clarify what the $32M covers. According to MindStudio's analysis, this figure includes the complete process: experimentation, data pipeline, evaluations, post-training. Not just the final pre-training run. The "one-tenth the cost of the previous SOTA" estimate relies on publicly reported compute deals — it's a comparison of estimates, not audited accounting.

What $32M does allow:

• A serious pre-training: 12 trillion tokens, a volume comparable to that of 2024-2025 frontier models.
• A dense architecture: no MoE with 800B total parameters, of which only a subset activates. The dense choice simplifies deployment and fine-tuning.
• Targeted post-training: 1.4M rollouts focused on agentic coding, rather than millions of rollouts across all domains.

What $32M does not allow: catching up with the leaders on the hardest benchmarks. On SWE-Bench Pro v2.0 Hard, Ripple-65B scores 79.1% versus 86.3% for Inkling 2 Small. On DeepSWE v1.1, the gap is brutal: 44.3% versus 72.1%. Light post-training is the deliberate trade-off made in the name of efficiency.

This efficiency lesson is not an isolated case. CacheRL shows that a Qwen3-4B model can reach 92% accuracy in tool-calling with 100 times less compute than GPT-5. The underlying trend is clear: in specialized domains, massive compute becomes a luxury, not a necessity.

Benchmarks: 80.5 on Terminal-Bench 2.1, but read the footnotes

Direct answer: the 80.52 on Terminal-Bench 2.1 is plausible but self-declared, and the model's own sheet admits that scores drop when using the official harness in certain configurations.

The most honest point on the Hugging Face sheet—and it is rare enough to be highlighted—concerns the fragility of benchmarks. Ricursive acknowledges that its scores were "universally lower" with the official Terminus 2 harness at 128K with summarization. In other words: the 80.52 figure depends on the evaluation configuration. Change the harness, change the context, and the model can lose several points.

The Decoder highlights the fundamental problem: all the figures come from the vendor; none have been independently replicated. This is the industry standard—labs publish their own evaluations—but it imposes a discipline of reading:

• Terminal-Bench 2.1 at 80.52: the best score among US open-weight models, but with only a 0.4 point lead over Reflection Beam 501B. A 501B parameter model beaten by a 65B model is spectacular—or it's a measurement gap.
• SWE-bench Verified at 76.5: solid, within the range of good open-weight coding models.
• DeepSWE v1.1 at 44.3%: this is where the model shows its limits. Long and difficult agentic tasks remain the territory of larger and more trained models.

For teams comparing open-source coding models, the rigorous approach is to cross-reference multiple sources: Kimi K2.7-Code claims, for example, 30% fewer reasoning tokens and an advantage in tool use—another vendor figure to be verified in the same manner.

How to deploy it yourself: DGX Spark, SGLang and formats

Direct answer: the NVFP4 repo fits on a single DGX Spark, and the BF16 repo runs with SGLang on 2x B200. It's one of the easiest open-weight models to self-host in its category.

Ricursive's dense choice isn't just a research decision: it's a deployment decision. Giant MoEs of 500 to 800B parameters require multi-node clusters; a 65B dense model in BF16 fits on two B200s. Concretely:

Ripple-65B also ships with DFlash2, a drafter for speculative decoding. The principle: a small model proposes tokens ahead of time, and the large model validates them in batches. On agentic coding workloads — lots of generated tokens, long sequences — the latency gain can be significant.

For those who don't want to manage GPUs, hosted inference is offered at $0.20 per million input tokens and $0.60 per million output tokens (November 2026, according to eesel AI). For comparison, that's well below the pricing of closed frontier models. For a coding agent consuming millions of tokens per day, the savings quickly become material.

If you're building a product around the model, also plan for the application layer: a landing page, documentation, a monitoring dashboard. Standard web hosting like Hostinger works perfectly fine for that part — the GPU, meanwhile, stays elsewhere.

How to verify the numbers yourself: the 4-step method

Direct answer: don't trust the vendor — reproduce Terminal-Bench 2.1 and SWE-bench Verified with the official harness, and compare against your own tasks.

Ricursive deserves credit for publishing the checkpoints and admitting the limits of its benchmarks. Take advantage of that. Here's the method I recommend to any serious team:

1. Download the weights from Hugging Face (ricursiveAI/Ripple-65B-Preview). NVFP4 format for a DGX Spark, BF16 for multi-GPU.
2. Reproduce with the official Terminus 2 harness, at 128K with summarization — the configuration where Ricursive admits the scores are lower. That's your honest baseline number.
3. Test on your own tasks. Agentic benchmarks measure generic terminal tasks. Your codebase, your workflows, your tools: that's what counts. Put together a set of 20-30 real tasks and measure the success rate.
4. Measure the real cost. Tokens consumed, latency, retry rate. A model that succeeds on 5% fewer tasks but costs 3x less may be the right choice.

This verification discipline is all the more important given that the current generation of open-weight models is innovating along very different axes: Sumi, for example, calls autoregression itself into question with uniform diffusion. When architectures change, historical benchmarks do an increasingly poor job of measuring what actually matters.

Where Ripple-65B Stands Against the Competition

Direct answer: it is the best quality/price/deployability ratio of American open-weight agentic coding, but not the best coding model outright.

Let's position it honestly. In the open-weight segment, Ripple-65B is the most performant US model on Terminal-Bench 2.1, beating models five to twelve times larger. In the closed segment, frontier models like GPT-5.5 (98.2) or Claude Opus 4.7 (94.3) remain ahead on the most complex agentic tasks—but at prices and with dependencies that are not suitable for all uses.

Ricursive's positioning is revealing, by the way. The lab, co-founded by Anna Goldie and Azalia Mirhoseini, specializes in AI for chip design. Ripple-65B is "a generalist variant of the proprietary model that powers our chip design platform." In other words: the model is not an end in itself, it is a demonstration of capability—and a signal sent to the market about what an efficient lab can produce.

Who is it the right choice for? Three profiles:

• Startups building coding agents who want an Apache 2.0, finetunable model with controlled inference costs.
• Regulated companies that cannot send their code to a third-party API.
• Researchers who want a 65B dense model, easy to fine-tune, with open checkpoints at every stage.

Who is it not the right choice for? If you need the best possible model for long agentic tasks, without license or budget constraints, closed frontier models remain ahead—and giant open-weight models like Inkling 2 Small dominate the toughest benchmarks.

❌ Common Mistakes

Mistake 1: confusing "open-weight" and "open-source"

An open-weight model publishes the weights, but not necessarily the training data, the pipeline code, or the license. Ripple-65B checks the license box (Apache 2.0), but the pre-training data is not published. Always verify what is actually open before talking about "open-source" in your communications.

Mistake 2: citing 80.52 on Terminal-Bench without specifying the configuration

The figure depends on the harness used. With the official Terminus 2 harness at 128K with summarization, the scores are lower — Ricursive itself acknowledges this. If you're comparing models, use the same configuration for all of them, otherwise the comparison is invalid.

Mistake 3: underestimating the inference cost of self-hosting

"Free, Apache 2.0" doesn't mean "free to run". Two B200s in BF16 represent a significant hardware investment or GPU rental. Do the math: for low volumes, hosted inference at $0.20/$0.60 per million tokens is almost always more economical. Self-hosting becomes cost-effective at high, consistent volumes.

Mistake 4: taking the vendor's numbers at face value

All of Ripple-65B's numbers come from the vendor and none have been independently replicated (The Decoder). This is true of almost every model, but it means you should test on your own tasks before any production commitment.

❓ Frequently Asked Questions

Can Ripple-65B Preview be used commercially?

Yes. The Apache 2.0 license covers all published checkpoints: commercial use, modification, distribution, and integration into paid products are allowed without restriction or royalties. It's one of the few agentic coding models of this level to be this permissive.

What are the $32M of training really worth?

The figure covers the full process — experimentation, data pipeline, evaluations, post-training — not just the final run. But it relies on the lab's claims and comparisons with public estimates. It's a credible order of magnitude, not an audited accounting.

Does Ripple-65B beat closed models?

No, not on the hardest tasks. On SWE-Bench Pro v2.0 Hard, it scores 79.1% versus 86.3% for Inkling 2 Small, and on DeepSWE v1.1, 44.3% versus 72.1%. On the other hand, it is very competitive on Terminal-Bench 2.1 among open-weight models, and its inference cost changes the economic equation.

What hardware is needed to run it?

The NVFP4 repo fits on a single DGX Spark. The BF16 repo runs with SGLang on 2x B200. That's significantly more accessible than the giant 500-800B parameter MoEs, which require multi-node clusters.

How does it differ from other recent open-weight models?

The dense 65B architecture, optimized for Nvidia hardware. While the open-weight competition pushes toward giant MoEs or hybrid architectures, Ripple-65B bets on deployment simplicity and specialization in agentic coding.

✅ Conclusion

Ripple-65B Preview is not the best coding model in the world — but it may be the best coding model that you can own: Apache 2.0, 65B dense parameters, deployable on reasonable hardware, at a training cost that rewrites the rules of the game. Don't trust the vendor's numbers: download the weights, reproduce the benchmarks with the official harness, and test it on your own tasks. That's the only verdict that counts.