Ripple-65B Beats a 501B Model at Coding, 80.5 to 80.1. On Terminal-Bench 4.0, That Model Scores 4.04%.

What Ricursive released on November 4

Ricursive Intelligence published Ripple-65B Preview on Hugging Face under the Apache 2.0 license. The model card describes a 65B-parameter dense Transformer, not a mixture of experts, pretrained from scratch on 12 trillion tokens of “predominantly-English and code” text, with support for up to 1M tokens of context.

Dense matters here. Every one of the 65B parameters runs on every token. The 4 models Ripple compares itself to are mixture-of-experts models with 501B to 975B total parameters, but only 23B to 55B of them active per token.

The release includes an instruction-tuned model in BF16, an NVFP4 checkpoint and a BF16 base model. Ricursive says the model “can fit on consumer hardware such as the NVIDIA DGX Spark,” and reports 250 tokens per second on 2 B200 GPUs at batch size 8, “prior to quantization or custom kernels.”

The cost claim is the other headline. Ricursive’s launch blog says “we developed and trained Ripple-65B Preview for only $32M,” covering the full R&D from scratch, and estimates that at “one-tenth the cost of the previous open-weight SOTA on agentic coding.” It also says its RL stage used 1.4M rollouts, “roughly 70x fewer rollouts than Reflection Beam.” The model card says the team “scaled from 1 to 5 people during development.”

Ripple is live on OpenRouter through GMICloud at $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05. Ricursive also lists Together AI and GMI Cloud as hosts.

The 4 rivals, on 2 versions of Terminal-Bench

Here is Ripple’s comparison set, with the scores from Ricursive’s table next to the same models on Vals.ai’s Terminal-Bench 4.0 leaderboard and their OpenRouter list prices.

TB = Terminal-Bench. TB 2.1 columns from the Ripple-65B model card; Ricursive marks some rival scores as its own in-house runs. TB 4.0 from Vals.ai, updated November 4, 2026. Prices from OpenRouter on November 5, 2026; Nemotron 3 Ultra also has a free variant.

On Ricursive’s numbers, Ripple wins by 0.42 points over Beam and by 16.7 points over Inkling, a model 15x its total size. On Vals.ai’s board, the same 3 rivals that appear there solve between 0.51% and 6.57% of tasks.

Why this isn’t a contradiction

It would be easy to read that table as “Ricursive inflated the numbers.” The sources don’t support that. Terminal-Bench 2.1 and Terminal-Bench 4.0 are different task sets. Vals.ai describes 4.0 as “a fresh set of 66 community-contributed, maintainer-reviewed tasks,” run with mini-swe-agent, a minimal harness that gives the model a single bash tool, an 8-hour limit per task and no step or cost limit.

Ricursive is upfront that harnesses move scores. Its card says that where no harness is named, it used OpenCode, and that Terminal-Bench 2.1 scores “were universally lower (with the exception of Muse Spark Open)” on the official Terminus 2 harness. You can see it in the table: Beam drops from 80.1 to 70.8 on Terminus 2, and Inkling from 63.8 to 56.2. Ripple holds at 80.5 on both.

So the honest read is narrower. Ripple beat its chosen rivals on a test where all 4 of them score above 55. On the newest independent run, those rivals collapse, and Ripple hasn’t been measured. Beating Beam on 2.1 tells you very little about where Ripple will land on 4.0.

The models the chart leaves out

Ricursive frames its claim as “US open-weight models,” and the comparison sticks to that scope. For a developer picking a model, nationality isn’t the filter; score and price are. On Vals.ai’s board, the open-weight models that matter sit far above Ripple’s rivals:

Kimi K3.1 Pro lands 1.51 points behind Claude Opus 5.5 at about a third of the cost per test. That’s the bar an open coding model has to clear today, and it’s the one Ripple’s launch table doesn’t show.

Price is where Ripple has a clear edge. At $0.20 / $0.60, it costs 10x less per input token than Kimi K3.1 Pro’s $2 and about 17x less per output token than Kimi’s $10. It’s also below every paid list price in its own chart. If its Terminal-Bench 4.0 score lands anywhere near the leaders, the cost per solved task could be very low. That’s an “if” until someone runs it.

Where Ripple wins and loses on its own card

Ripple’s table isn’t a clean sweep, and Ricursive doesn’t pretend it is. Among the 5 models it compares:

It leads on both Terminal-Bench 2.1 rows (80.5) and on SWE-bench Pro v2.0 Hard, at 79.1 to Beam’s 76.5 and Inkling’s 2.0.

It trails on SWE-bench Verified (76.5 to Beam’s 81.6), SWE-bench Pro v1.0 (61.42 to Beam’s 66.1), DeepSWE 1.1 (44.3 to Muse Spark Open’s 60.0), SciCode (39.4 to 58.5) and MCP Atlas (71.67 to 89.2). On MMLU and GPQA Diamond, it scores lowest of the 5.

Ricursive also calls this “relatively light posttraining, with a focus on agentic coding.” That fits the pattern: strong on terminal tasks, weaker on tool-heavy and science work.

What this means if you code with AI

Treat “beats a model 10x its size” as a claim about the rivals. A win is only as strong as the models it beats. Here, the 4 rivals score 0.51% to 6.57% on the newest independent run of Terminal-Bench. Always check who’s missing from a launch chart.

Wait for the independent number, or run your own. Ripple isn’t on Vals.ai’s Terminal-Bench 4.0 board yet. At $0.20 / $0.60, a test drive on 10 or 20 of your real tasks costs very little. Run it in the agent harness you actually use, since Ricursive’s own card shows the harness can move a score by 9 points.

The local story is the real news. An Apache 2.0 coding model with 1M context that fits on a DGX Spark is useful whether or not it tops a chart. For private code, offline work or flat costs, a strong local model changes what “free AI code generator” can mean.

For hosted agent work today, the board hasn’t moved. Kimi K3.1 Pro is the open-weight model to beat at 63.64%, and GPT-6.1 Sol remains the cheap closed option at 55.05% for $1.72 a test.

Full disclosure: we sell an AI coding agent. MegaBrain’s free plan needs no card and includes the free models, plus your own API keys and local models served by Ollama or LM Studio, with no limits from us. See the free AI code generator or the plans. MegaBrain is not affiliated with Ricursive Intelligence or Vals.ai.

Fun facts

Ripple’s model card cites an Olmo 3 baseline as its starting point and says its final recipe improves compute efficiency on code by 16x over the Olmo 3 ladder, rising to 55x for the full run.

Ricursive’s launch blog footnotes its “one-tenth the cost” estimate with a report that Reflection committed $150M per month for compute, and that Beam took 4 weeks to pretrain and 4 weeks to post-train.

The longest training stages ran at a 2M-token sequence length, twice the 1M context the model card advertises.

Sources

All pages opened November 5, 2026.

• Hugging Face — Ripple-65B Preview model card (specs, benchmark table, harness notes, Apache 2.0)
• Ricursive Intelligence — Ripple launch blog ($32M R&D cost, 1.4M RL rollouts, availability), November 4, 2026
• Vals.ai — Terminal-Bench 4.0 leaderboard (66 tasks, mini-swe-agent harness, updated November 4, 2026)
• OpenRouter — Ripple-65B Preview model page and endpoints (GMICloud, BF16, 1M context)
• OpenRouter — Models API (list prices for Beam 1.0, Muse Spark Open, Inkling, Nemotron 3 Ultra, Kimi K3.1 Pro)