Ripple-65B Preview Explained: 80.5 on Terminal-Bench, $32M Build Cost and the 65B Model Taking On 501B Rivals

Illustrative image: NVIDIA DGX Spark — the desktop AI computer Ricursive says Ripple-65B Preview's NVFP4 build is designed to run on. This is the hardware the model targets, not the model itself: a model has no product photo. Photo: Daniel Lu, CC BY-SA 4.0, via Wikimedia Commons.

TL;DR — Ripple-65B Preview explained in one line: a little-known chip-design lab, Ricursive Intelligence, has released a 65-billion-parameter open-weight AI model that it says beats 501-billion-parameter rivals at agentic coding — scoring 80.5 on Terminal-Bench 2.1 after a $32 million build. The catch: every benchmark number so far is the company's own, and no independent test has confirmed them yet.

📌 KEY HIGHLIGHTS

• What it is: Ripple-65B Preview is a 65B dense Transformer from Ricursive Intelligence, pretrained from scratch on 12 trillion tokens, with up to 1M tokens of context — released 4 November 2026 under the Apache 2.0 licence.
• The headline number: Ricursive reports 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4) and Thinking Machines Inkling 975B (63.8).
• The cost claim: the whole project cost $32 million, which Ricursive calls one-tenth the cost of the previous open-weight leader on agentic coding.
• Why it runs on one box: because it is dense, not mixture-of-experts, an experimental NVFP4 build is designed for a single NVIDIA DGX Spark desktop computer.
• The honest caveat: all benchmark scores are vendor-reported. No independent evaluation had appeared as of 7 November, and the comparison covers US open-weight models only.

Ripple-65B Preview explained: what happened

Ricursive Intelligence is not a name most AI watchers knew a week ago. The Palo Alto lab, co-founded by former Google researchers Anna Goldie and Azalia Mirhoseini, describes itself as a full-stack AI lab building self-improving systems by co-optimising hardware, software and model — its day job is using AI to design chips. To power its chip-design platform, it trained a language model from scratch. On 4 November 2026, it released a general-purpose variant of that internal model as Ripple-65B Preview.

The release is small by 2026 standards and that is the point. Ripple has 65 billion parameters — 64.8 billion, to be exact — and every one of them is active on every token. Most frontier open models today are mixture-of-experts designs, where only a fraction of the parameters switch on per token: Reflection Beam has 501 billion total parameters, Thinking Machines' Inkling, a 975-billion-parameter model, activates far fewer. Ripple is dense, thin and deep: 80 layers with a narrow 5,120-wide model dimension. Ricursive says that shape is what lets it run efficiently on local hardware instead of a data centre.

The weights are free. Ripple ships under Apache 2.0, so anyone — including commercial users — can download, modify and deploy it. Hugging Face hosts four checkpoints: the instruction-tuned BF16 model (about 130 GB), an experimental NVFP4 build aimed at local hardware, a pretrain-only base model and a DFlash2 draft model. Hosted APIs are available through Together AI, GMI Cloud, OpenRouter and Fireworks.

Ripple-65B Preview specs and benchmarks

The specification sheet, from Ricursive's model card:

Specification	Ripple-65B Preview
Parameters	64.8B, dense — all active on every token
Architecture	Dense Transformer, 80 layers, d_model 5,120
Training data	12 trillion tokens, mostly English text and code
Context length	Up to 1,048,576 tokens (claimed; see the hosted cap below)
Knowledge cutoff	17 December 2025
Licence	Apache 2.0
Checkpoints	BF16 instruct, NVFP4 (experimental), BF16 base, DFlash2
Build cost	$32 million full R&D cost, per Ricursive

The benchmark table is where the story gets interesting — and where it needs care. These are Ricursive's own numbers, from its model card, comparing only US open-weight models. Chinese open models such as DeepSeek, Qwen and Kimi are not in the table at all.

Benchmark (higher is better)	Ripple-65B	Beam 501B	Muse Spark Open 816B	Inkling 975B	Nemotron 3 Ultra 550B
Terminal-Bench 2.1 (self-reported)	80.5	80.1	76.4	63.8	56.4
Terminal-Bench 2.1 (Terminus 2 harness)	80.5	70.8	78.7	56.2	55.1
SWE-bench Pro v2.0 Hard	79.1	76.5	72.4	39.6	30.7
SWE-bench Verified	76.5	81.6	78.7	77.6	71.9

Read the table as a whole, not just the headline row. Ripple wins the terminal-agentic test by 0.4 points over Beam — a hair, not a rout — and wins the hard bug-fixing test, SWE-bench Pro v2.0 Hard, more clearly. But on everyday bug-fixing, SWE-bench Verified, Beam is still ahead at 81.6 against Ripple's 76.5. Ripple also trails on DeepSWE 1.1 (44.3), SciCode (39.4, the lowest score in its own table) and MCP Atlas (71.67). On maths it is strong: 98.3 on AIME 2025 and 90.5 on HMMT February 2026. And Ricursive's own card carries a warning that matters: the scores are hard to compare because some rival figures come from each lab's own model card and some from Ricursive's in-house runs, on different harnesses. With the official Terminus 2 harness, Beam's Terminal-Bench score falls from 80.1 to 70.8 while Ripple's stays at 80.5 — which flatters Ripple, and is exactly why independent results are needed before anyone calls this settled.

Illustrative image: the rear of an NVIDIA DGX Spark, showing its ConnectX-7 networking port. Ripple's small parameter count is what makes it a fit for a single desktop box like this. Photo: Daniel Lu, CC BY-SA 4.0, via Wikimedia Commons.

Why a 65B model beating 501B rivals matters

The result is surprising because of the size gap. Ripple is about 15 times smaller than Inkling (975B total parameters) and roughly 8 times smaller than Beam (501B), yet it matches or beats them on the two agentic-coding tests Ricursive chose to lead with. Two things explain how, according to the launch post.

First, the training was cheap by frontier standards. Ricursive says it developed and trained Ripple for $32 million all-in — pretraining-ladder experiments, data mixing, evals, debugging, synthetic-data generation, post-training and the final runs — and calls that one-tenth the cost of the previous open-weight leader on agentic coding. Its footnote contrasts that with reporting that Reflection committed $150 million a month for compute. The reinforcement-learning climb was also frugal: Ricursive says it reached 80.5 on Terminal-Bench with only 1.4 million rollouts, roughly 70 times fewer than Reflection Beam needed.

Second, the recipe was borrowed and refined rather than invented. Ricursive started from the Allen Institute for AI's open Olmo 3 training ladder and iterated on it at small scale — testing interventions, rejecting the ones that did not scale, and ending with a ladder it says is 16 times more compute-efficient on code at 4-billion-parameter scale. The architecture adds gated attention, global NoPE (no positional embeddings), value residuals and FFN activation clipping, and the team held 40% model FLOPs utilisation through training with simple sharding, no custom kernels and no quantisation.

That efficiency is the real story for the open-model world: if a small lab can reach the top of a respected coding benchmark for $32 million, the cost of building a competitive open model just fell sharply. The "Preview" label is the counterweight — post-training was relatively light and focused on agentic coding, so this is a first cut, and the weak spots in its own table (SciCode, DeepSWE, tool use) show where that lightness shows.

What it costs and how to run it

The weights are free under Apache 2.0, so the only cost is compute. Hosted, Ripple costs $0.20 per million input tokens and $0.60 per million output tokens on OpenRouter (served by GMI Cloud), with cached input at $0.05 — Together AI lists cached input slightly cheaper, at $0.02. At roughly ₹90 to the dollar, that is about ₹18 per million input tokens in: pocket change next to frontier-model API prices, and cheap enough to run a coding agent all day without watching the meter.

Self-hosting is the more interesting option. The BF16 instruction-tuned model is about 130 GB, which needs data-centre GPUs. The experimental NVFP4 build is the one aimed at local hardware: Ricursive's launch post names the NVIDIA DGX Spark — a desktop AI computer with 128 GB of unified memory — as the kind of machine Ripple is designed to run on. That is why the photos in this post show a DGX Spark.

One limit to know before you switch anything over: the model card claims up to 1 million tokens of context, but OpenRouter's live listing caps context at 16,384 tokens. Through the hosted API today, Ripple is far more short-sighted than the model card suggests — a real constraint for the long agentic-coding sessions it was built for. Self-hosting removes that cap, but then you need the hardware.

Why it matters for India readers

There is no India-specific launch, price or availability for Ripple-65B Preview — it is a global, free download, and this post does not invent an Indian angle for it. The relevance is practical. Indian developers, students and startups can download Apache 2.0 weights free from Hugging Face and use them commercially, with no licence fee and no vendor lock-in — the kind of open model that lowers the entry cost of building coding agents. The hosted API, at about ₹18 per million input tokens, is cheap enough for hobby projects and early-stage products. And the local-running story matters to India's growing community of builders who run models on their own machines: a 65B model that fits on a desktop-class box is a different proposition from a 500B-plus model that needs a cluster. The honest next step for anyone interested is the model card itself, not the headline number.

Should you care?

✅ Worth trying if you build coding agents and want a small, cheap, commercially-usable open model you can self-host — or if you run models locally and want to test the DGX Spark claim yourself.

⏳ Wait if you need the best bug-fixing scores (Beam still leads SWE-bench Verified), long hosted context (the 16K API cap bites), or verified benchmarks — no independent evaluation had landed as of 7 November, so treat the headline as a claim, not a verdict.

FAQs

What is Ripple-65B Preview?
Ripple-65B Preview is a 65-billion-parameter open-weight language model released on 4 November 2026 by Ricursive Intelligence, a chip-design lab. It is a dense Transformer trained from scratch on 12 trillion tokens, tuned for agentic coding, and released under the Apache 2.0 licence.

Is Ripple-65B Preview free to use?
Yes. The weights are free to download from Hugging Face under Apache 2.0, which permits commercial use. You only pay for compute — either your own hardware or a hosted API, which costs $0.20 per million input tokens and $0.60 per million output tokens on OpenRouter.

How good is Ripple-65B at coding?
On Ricursive's own numbers it is very good at terminal-based agentic work, scoring 80.5 on Terminal-Bench 2.1, and strong on hard bug-fixing (79.1 on SWE-bench Pro v2.0 Hard). But it trails bigger rivals on everyday bug-fixing (76.5 on SWE-bench Verified, against Beam's 81.6) and on science and tool-use tests. All figures are vendor-reported; no independent evaluation had appeared as of 7 November.

Can I run Ripple-65B on my own computer?
The full BF16 model is about 130 GB and needs data-centre GPUs, but the experimental NVFP4 build is designed for a single NVIDIA DGX Spark desktop computer with 128 GB of unified memory. Note the hosted API caps context at 16,384 tokens even though the model card claims up to 1 million — self-hosting is the way to use the long context.

Takeaway: Ripple-65B Preview is the most efficient open-model result of the year — a $32 million, 65B model matching 500B-plus rivals on agentic coding. The numbers are the vendor's own, so the headline is a claim to test, not a verdict — but if independent results confirm even half of it, the cost of building a competitive open coding model has just fallen sharply.

If you build coding agents, the honest next step is the model card itself: read Ricursive's benchmark table and its harness caveats on Hugging Face before you switch anything over — and compare it with Mistral Large 4, another recent open-weights release, and NVIDIA's Cosmos 3 Nano open-weight models to see how fast this space is moving.

Sources

• Ricursive Intelligence — "Introducing Ripple-65B Preview, a Frontier Open-Weight Model" (launch blog, 4 Nov 2026) — ricursive.com/ripple
• Hugging Face — Ripple-65B Preview model card — huggingface.co/ricursiveAI/Ripple-65B-Preview
• VentureBeat — "Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks" (5 Nov 2026) — venturebeat.com
• DepthGrid — "Ripple-65B Preview: 65B Dense Open Model, Benchmarks, Apache 2.0" (6 Nov 2026) — depthgrid.in
• eesel.ai — "Ripple-65B Preview: benchmarks, price and how to run it" (Nov 2026) — eesel.ai