Cognition, the maker of the Devin coding agent, launched SWE-2 Flash on 4 November and posted on X at 19:55 UTC that “On leading evals, it outperforms all American open-weight models, even ones 3x larger”. About 91 minutes earlier by our arithmetic from the two posts’ times, Ricursive Intelligence had posted that Ripple-65B Preview “sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks”. On Terminal-Bench 2.1, a test both labs report, Cognition lists SWE-2 Flash at 89.2 and Ricursive lists Ripple at 80.5: a gap of 8.7 points, by our arithmetic, across two self-reported tables.
The two releases differ on access. Ripple’s weights are public on Hugging Face under Apache 2.0. Cognition says it is “not sharing the weights publicly yet”, citing cyber risk, and is building a program to share them with partners and researchers.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on the difference between a model and the harness around it. It teaches the general idea only and does not discuss Cognition or Ricursive.
Scores on the three coding evals both labs report, from Cognition’s SWE-2 Flash blog and Ricursive’s Ripple-65B Preview model card (4 November 2026). Self-reported by each lab, by different teams and harnesses; Cognition does not state its harness. Gaps are our arithmetic. Chart: FourWeekMBA.
The key insight: On the Terminal-Bench 2.1 coding test, Cognition lists SWE-2 Flash at 89.2 and Ricursive lists Ripple-65B Preview at 80.5, a gap of 8.7 points by our arithmetic, but the two scores come from two self-reported tables, and only Ripple’s weights are public.
Cognition’s launch blog calls SWE-2 Flash “a fully American coding model that sets a new bar for intelligence at low cost and high speed”. It is built on Inkling-Small from Thinking Machines, which the blog describes as “a model with 276B total and 12B active parameters”.
Cognition says a mid-training stage added before reinforcement learning lifted the base model on Terminal-Bench 4.0 from 0.9% to 20.0% and on FrontierCode 2.0 from 3.2% to 24.0%. The model is available in Devin Desktop and CLI “at one-fifth the price and 3x the speed of its much larger sibling, SWE-2”. A second Cognition post says it is free until the end of the month for Pro, Max and Teams users.
Terminal-Bench 2.1 scores as each lab reports them: SWE-2 Flash from Cognition’s launch blog and Ripple-65B Preview from Ricursive’s model card, both 4 November 2026. Reflection Beam and Muse Spark Open carry the same score in both tables. All figures are self-reported and not independently verified.
What Was Posted on X
Introducing SWE-2 Flash, our new fast & efficient coding model.
On leading evals, it outperforms all American open-weight models, even ones 3x larger.
To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers.
Cognition’s launch post for SWE-2 Flash, 19:55 UTC on 4 November. It says the model outperforms all American open-weight models on leading evals and that Cognition is building a program to share the weights securely with partners and researchers.
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
— Ricursive Intelligence (@RicursiveAI) Nov 04 18:24
Ricursive Intelligence’s launch post for Ripple-65B Preview, 18:24 UTC the same day. It says Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks.
Really excited to release SWE-2 Flash. Our latest model, based on @thinkymachines Inkling-Small and optimized for real-world coding tasks. Let us know what you think!
Scott Wu (@ScottWu46) posted this at 05:01 UTC on 5 November 2026. He calls SWE-2 Flash Cognition’s latest model, says it is based on Thinking Machines’ Inkling-Small and optimized for real-world coding tasks, and asks users for feedback.
The Same Three Tests, Two Tables
Three evals appear in both labs’ tables. On Terminal-Bench 2.1, Cognition lists SWE-2 Flash at 89.2 and Ricursive lists Ripple at 80.5. On DeepSWE 1.1 the figures are 60.1 and 44.3, and on SWE-bench Pro v2 Hard 89.9 and 79.1: gaps of 15.8 and 10.8 points, by our arithmetic.
Some reference scores match across the two tables. Both list Reflection’s Beam at 80.1 on Terminal-Bench 2.1, 43.3 on DeepSWE 1.1 and 76.5 on SWE-bench Pro v2 Hard, and both list Meta’s Muse Spark Open at 76.4 and 60.0 on the first two. On SWE-bench Pro v2 Hard they diverge: Cognition lists Muse Spark Open at 82.4, while Ricursive’s in-house run puts it at 52.3, a 30.1-point difference for the same model, by our arithmetic.
Cognition’s page does not state the harness it used. Ricursive’s card says it uses Terminus 2 for Ripple on SWE-bench Pro v2.0 Hard and DeepSWE 1.1, and that “Replicating benchmark scores has been challenging, particularly for modern agentic evals”. Ripple does not appear in Cognition’s table.
Size, Speed and the Full Table
SWE-2 Flash has about 4.2 times Ripple’s 65 billion total parameters, by our arithmetic, but it is a mixture-of-experts model with 12 billion active; Ripple is dense, so it uses about 5.4 times as many parameters per token.
In Cognition’s own eight-row table, SWE-2 Flash scores above Beam 1.0 in all 8 rows and above Muse Spark Open in 6 of the 7 rows where Muse has a score, by our count. Muse Spark Open leads on SWE-Atlas QA (54.0 vs 50.0), and on DeepSWE 1.1 the lead is 0.1 points. The same table lists Claude Haiku 5.5 above SWE-2 Flash in 7 of the 8 rows, and DeepSeek V4.1 Flash above it in 4, including Terminal-Bench 2.1 (91.0).
Why Cognition Is Holding the Weights
The blog says Cognition targeted cybersecurity work in training, raising the model’s ExploitBench score from 11.3% to 34.8%. It adds: “These same capabilities can also be used by malicious attackers, as demonstrated by a recent surge in cyber attacks, so we are not sharing the weights publicly yet”. It says more details on the program will follow “in a few weeks”.
As we read the two launches, a lead over American open-weight models now comes in two forms: Ricursive released its weights under Apache 2.0, while Cognition compares SWE-2 Flash with open-weight models but is not sharing its weights publicly yet.
As we read the tables, the reference scores are the shared ground: both list Reflection’s Beam at 80.1 on Terminal-Bench 2.1, while they disagree by 30.1 points on Muse Spark Open’s SWE-bench Pro v2 Hard score, by our arithmetic.
Cognition, Introducing SWE-2 Flash
“These same capabilities can also be used by malicious attackers, as demonstrated by a recent surge in cyber attacks, so we are not sharing the weights publicly yet”
Three Implications
FOR DEVIN USERS Cognition says SWE-2 Flash runs at one-fifth the price and 3x the speed of SWE-2, and a Cognition post says it is free to month end for Pro, Max and Teams users.
FOR BENCHMARK READERS Cognition does not state its harness, and Ricursive says it uses Terminus 2 for Ripple on two of the three shared tests, so the 8.7-point gap compares tables, not one controlled run.
FOR TEAMS THAT NEED WEIGHTS Ripple-65B Preview can be downloaded today under Apache 2.0; Cognition says SWE-2 Flash weights will go to trusted partners and research labs through a program it is still building.
The framework says: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, the two labs drew the line in different places on the same day: Ricursive opened Ripple’s weights, while Cognition keeps SWE-2 Flash’s weights inside Devin Desktop and CLI and plans gated access through a program it is still building. The framework does not discuss Cognition or Ricursive. For the full playbook, read The Open vs Closed Meta-Framework on businessengineer.ai.
All scores are self-reported by each lab; we found no independent evaluation of SWE-2 Flash or Ripple in what we read. The cross-lab gaps compare two tables run by different teams, with Cognition’s harness unstated, so they are not one controlled test. Cognition has not said which partners will get the weights, or on what terms.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Two labs claimed a lead over American open-weight coding models on 4 November, 91 minutes apart by our arithmetic. On Terminal-Bench 2.1, Cognition lists SWE-2 Flash at 89.2 and Ricursive lists Ripple at 80.5, by their own tables. Ripple’s weights are public; SWE-2 Flash’s are not yet.
A note on sourcing. SWE-2 Flash details, scores and quotations are from Cognition’s launch blog and X posts; Ripple-65B Preview details are from Ricursive Intelligence’s Hugging Face model card and X post, all read on 4 November 2026. All benchmark scores are self-reported by each lab; Cognition does not state its harness, and some rival scores in Ricursive’s table are its in-house runs. Gaps, ratios and counts are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
