AI news ·
Cognition launches SWE-2 Flash coding agent with triple the speed and one-fifth the cost of its predecessor
Cognition's SWE-2 Flash runs at 200 tokens per second, three times faster than SWE-2 and at one-fifth the price. The 276B-parameter model targets routine coding tasks but won't release its weights publicly due to cybersecurity concerns.
Cognition released SWE-2 Flash on November 4, 2026, a coding model that runs at three times the speed and one-fifth the price of its larger SWE-2 sibling. The model is designed for routine software-agent tasks such as automations, code review, feature development, and documentation updates, rather than chasing uniform benchmark leadership.
Flash is available in Devin Desktop and CLI, with free use offered through November 30, 2026. It has 276 billion total parameters and 12 billion active parameters, matching Thinking Machines Lab's Inkling-Small, the mixture-of-experts base model Cognition adapted. Thinking Machines released Inkling-Small's full weights on July 30, 2026.
Benchmark results show targeted trade-offs
Cognition reported evaluation scores that show substantial gains over Inkling-Small but mixed results against SWE-2. Flash scored 20.0 on Terminal-Bench 4.0 compared to SWE-2's 27.3 and Inkling-Small's 0.9. On FrontierCode 2.0, Flash edged out SWE-2 with a score of 24.0 versus 23.8. SWE-2 held the lead on DeepSWE 1.1 at 73.0 to Flash's 60.1.
The results position Flash as a cost-and-speed alternative rather than a wholesale replacement. Cognition said Flash outperforms all American open-weight models on these evaluations, including models three times its size. The company illustrated generation speed by replaying each model's Terminal-Bench 4.0 gsea-proteomics run at simulated rates of 200 tokens per second for Flash and 60 for SWE-2, with transcripts shortened to equal length.
Mid-training and reinforcement learning recipe
Cognition said this is the first time its recipe placed a large-scale mid-training stage before reinforcement learning. The team curated a held-out coding dataset meant to represent real software-engineering scenarios and compared candidate data mixtures by how many tokens each required to reach the baseline run's validation loss. Token efficiency became a selection criterion.
The final mixture contained a few trillion tokens drawn from scientific articles, blogs, code, and other sources. Training proceeded in three stages so the scarcest, highest-quality data appeared later. Cognition reported that held-out validation loss improved at a stable log-linear rate across those stages.
Large-scale reinforcement learning followed. Cognition attributed most of the improvement over its SWE-2 recipe to higher-quality and more diverse tasks rather than recipe adjustments. The team audited task instructions, environment setup, and verifiers, and added task types intended to build broader reasoning and problem-solving for real Devin workflows.
Restricted weight access and security concerns
Flash's weights are not being shared publicly. Cognition is building a secure sharing program for trusted industry partners and research labs. Approved partners would be able to fine-tune the model and run it in customer-controlled environments, including on-premise infrastructure, while researchers could study behavior and potential misuse.
Cognition reported that targeted cybersecurity training raised ExploitBench from Inkling-Small's 11.3% to 34.8%, but warned that the same capabilities could assist malicious attackers. That concern prompted the decision to withhold public weights. A Cognition security post from October 26, 2026 described AI as speeding up work for both attackers and defenders and discussed PHP and Node.js remote-code-execution vulnerabilities disclosed on October 21, 2026.
Why this matters for product development teams
Flash changes the economics of AI-assisted coding for everyday tasks. A model that runs at 200 tokens per second and costs one-fifth the price of its larger sibling makes it practical to integrate coding agents into tighter CI/CD loops, automated code review pipelines, and documentation workflows without blowing out infrastructure budgets. For teams already using Devin or evaluating AI agent courses, the speed improvement directly reduces wait times on routine pull requests and feature scaffolding. The restricted weights, however, mean enterprises that need air-gapped or on-premise deployment will have to wait for Cognition's partner program to materialize before they can run Flash in sensitive environments.