Two coding models claim the US open-weight lead: Ripple-65B ships Apache 2.0 weights, SWE-2 Flash keeps them back

Two coding models arrived on 4 November, and both say they beat every American open-weight model. One of them is open weight. Ricursive Intelligence published Ripple-65B Preview on Hugging Face under Apache 2.0. Cognition launched SWE-2 Flash, trained on top of an open-weight base model, and is not releasing its weights publicly. Every score below comes from each vendor’s own tables.

Ripple-65B Preview: weights you can run

Ricursive describes Ripple-65B Preview as a 65B-parameter dense Transformer pretrained from scratch on 12 trillion tokens, with up to 1M tokens of context. The release has three checkpoints on Hugging Face, all tagged Apache 2.0 and not gated:

• BF16 instruct model (the main repo, with an Apache 2.0 LICENSE file).
• NVFP4 instruct model, labeled experimental. Ricursive calls it “a conservative quantization” that uses NVFP4 for the MLP layers and keeps attention in BF16.
• BF16 base model, a pretrain-only checkpoint.

A fourth repo holds a DFlash2 variant. Ricursive says the model fits on consumer hardware such as the NVIDIA DGX Spark, and that it is also served by Together AI, GMI Cloud and OpenRouter.

On its launch page, Ricursive says it developed and trained the model “for only $32M,” which it calls its “full R&D cost starting from scratch,” and estimates that is one-tenth the cost of the previous open-weight leader on agentic coding, “to the best of our knowledge.” The model card says the team grew from one to five people during development and started from an Olmo 3 baseline.

Its headline number is 80.5 on Terminal-Bench 2.1, against 80.1 for Reflection Beam 501B, 76.4 for Meta Muse Spark Open 816B, 63.8 for Thinking Machines Inkling 975B and 56.4 for NVIDIA Nemotron 3 Ultra 550B. The same table shows where Ripple trails: 39.4 on SciCode (Muse Spark Open: 58.5) and 44.3 on DeepSWE 1.1 (Muse Spark Open: 60.0). Ricursive says it filled gaps in rivals’ published scores with its own evals, ran its own model with OpenCode where no harness was specified (Terminus 2 for SWE-bench Pro v2.0 Hard and DeepSWE 1.1), and found Terminal-Bench 2.1 scores “universally lower” for most models under the official Terminus 2 setup, calling its results “a best-effort attempt at a fair comparison.”

SWE-2 Flash: weights held back

Cognition built SWE-2 Flash on Thinking Machines’ open-weight Inkling-Small (276B total, 12B active parameters), adding a mid-training stage before reinforcement learning. Cognition says that lifted the base from 0.9% to 20.0% on Terminal-Bench 4.0 and from 3.2% to 24.0% on FrontierCode 2.0, and that SWE-2 Flash “outperforms all American open-weight models, even models that are 3x larger.”

It runs in Devin Desktop and CLI at “one-fifth the price and 3x the speed” of SWE-2, and the post says it is free through 30 November. It gives no per-token price.

Cognition deliberately trained for cybersecurity work “to help defenders,” raising ExploitBench from 11.3% to 34.8%. Because attackers could use the same skills, “we are not sharing the weights publicly yet.” Instead it is building a program to share weights with trusted industry partners and research labs, with “more on this in a few weeks.” That follows Cognition’s recent focus on cost per task, which we covered when it tied extra Devin usage to an efficiency score.

Reading the two tables side by side

This is not a head-to-head test. Neither vendor lists the other’s model.

• Terminal-Bench 2.1: Cognition’s table gives SWE-2 Flash 89.2. Ricursive’s gives Ripple 80.5. Different harnesses and setups may sit behind each number, so the gap is not a measured comparison.
• Shared reference points: both tables list Reflection Beam (80.1) and Muse Spark Open (76.4) at the same Terminal-Bench 2.1 scores. Reflection announced Beam in October.
• Different Inklings: Cognition compares against Inkling-Small (276B, 64.7), Ricursive against the larger Inkling 975B (63.8).
• Why “American” matters: Cognition’s own table puts DeepSeek V4.1 Flash 1017 at 91.0 on Terminal-Bench 2.1, above SWE-2 Flash. DeepSeek posted those weights in October.

Why this matters

If you need a coding model on your own hardware, to fine-tune, audit or run offline, Ripple-65B Preview is the one of the two you can download today, under Apache 2.0, with an NVFP4 build Ricursive aims at DGX Spark-class machines. SWE-2 Flash is, for now, a Devin model you can try free until 30 November, with weights promised only to vetted partners. Both “best American open-weight” claims rest on self-reported tables that never test each other, and Cognition’s own table shows a Chinese open-weight model ahead, so run your own tasks before switching.