Cognition 发布 SWE-2 Flash 编码模型，速度提升三倍、价格降至五分之一

Cognition announced SWE-2 Flash on November 4, 2026, and reports that the coding model runs at three times the speed and one-fifth the price of its larger SWE-2 sibling. Available in Devin Desktop and CLI, it is aimed at routine software-agent work rather than uniform benchmark leadership.

The intended jobs include automations, code review, feature development, and documentation updates, with free use offered through November 30, 2026. Flash 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 Gains and Trade-Offs

Cognition reports the following evaluation scores, which show substantial gains over Inkling-Small but mixed results against SWE-2:

Terminal-Bench 4.0 0.9 20.0 27.3

FrontierCode 2.0 3.2 24.0 23.8

DeepSWE 1.1 17.4 60.1 73.0

In Cognition’s comparison, Flash edges SWE-2 on FrontierCode 2.0, while SWE-2 scores higher on Terminal-Bench 4.0 and DeepSWE 1.1. The results frame Flash as a cost-and-speed alternative rather than an across-the-board replacement for its larger sibling.

Cognition says Flash outperforms all American open-weight models on the leading evaluations it reports, including models three times larger. Its generation-speed illustration is specific about the method: each model’s own Terminal-Bench 4.0 gsea-proteomics run is replayed 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

Cognition says this is the first time its recipe has placed a large-scale mid-training stage before reinforcement learning. To choose the data mixture, the team curated a held-out coding dataset intended to represent real software-engineering scenarios. Candidate mixtures were compared by how many tokens each required to reach the baseline run’s validation loss, making token efficiency 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 reports that held-out validation loss improved at a stable log-linear rate across those stages.

Large-scale reinforcement learning followed. Cognition attributes most of the improvement over its SWE-2 recipe to higher-quality and more diverse tasks rather than to 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

Flash’s weights are not being shared publicly yet. Cognition is building a secure sharing program for trusted industry partners and research labs, with further details promised. The planned access would let approved partners fine-tune the model and run it in sensitive customer-controlled environments, including on-premise infrastructure, while researchers could study behavior and potential misuse.

Cognition reports that targeted cybersecurity training raised ExploitBench from Inkling-Small’s 11.3% to 34.8%, but warns that the same capabilities could assist malicious attackers, prompting the decision to withhold public weights.

Cognition’s October 26, 2026 security post describes AI as speeding up work for both attackers and defenders and discusses the PHP and Node.js remote-code-execution vulnerabilities disclosed on October 21, 2026.

The release separates immediate utility from broader access. Inside Devin, Flash is a reported cost-and-speed option for everyday engineering work. Beyond that product, its value to enterprises and researchers depends on whether the planned program can deliver fine-tuning and customer-controlled deployment under the security restrictions Cognition has set.