# Cognition SWE-2 Flash: Devin Access, Speed Claims and the Weight-Release Limits

SWE-2 Flash brings a smaller Inkling-based coding model to Devin. Its launch separates lower-cost agent work from public weights and verified performance.

By Paras Madan | 2026-11-05


Code review, documentation changes and routine fixes can consume a lot of model calls. A cheaper coding model is useful when it reduces that bill without moving the cost into failed attempts and developer cleanup.

Cognition's [November 4 announcement](https://cognition.com/blog/swe-2-flash) puts SWE-2 Flash in Devin Desktop and CLI, with free access through November 30. The company claims one-fifth the price and three times the speed of SWE-2.

## What the launch establishes

Flash builds on Inkling-Small, with 276 billion total parameters and 12 billion active per token, followed by additional training.

Cognition's benchmarks show tradeoffs: Terminal-Bench 4.0 scores are 20.0 for Flash and 27.3 for SWE-2, while FrontierCode 2.0 slightly favors Flash, 24.0 to 23.8. These are vendor results.

The speed animation replays shortened transcripts at simulated rates of 200 versus 60 tokens per second. It is not a live throughput measurement. The post also gives no standalone per-token price sheet.

Cognition is withholding public weights, citing cybersecurity misuse risks. A secure access program for trusted partners and research labs is being developed, with details promised in the coming weeks.

The distinction matters for buyers: trying a model inside an existing agent product and preparing to host its weights require different plans. The announced future program should not become a deployment dependency today.

## Test the whole engineering workflow

Teams evaluating Flash in [Devin Desktop](https://devin.ai/desktop) or the [Devin CLI](https://devin.ai/cli) should check their account's access and billing details. A relative price claim does not tell you how much a finished change will cost.

Start with a small, fixed set of recent tasks. Include straightforward changes, ambiguous bug reports and work that previously needed several attempts. Preserve the same repository state and acceptance criteria for each comparison.

Count successful changes, retries, review effort and total turnaround time. Keep the time spent generating tokens separate from tool execution, waiting for tests and human review. A faster response can still produce a slower task if it sends the workflow down the wrong path.

Then look at the failures. If a model handles most documentation changes well but struggles with a particular class of debugging, that is useful information for deciding which work to assign it. A single average score hides those differences.

## Keep access decisions separate

For teams that need on-premises deployment, weight access is a requirement to verify before investing in integration work. A hosted trial can teach you about output quality without settling questions about self-hosting, fine-tuning or distribution rights.

SWE-2 Flash offers another option for evaluating the cost of routine agent work. The strongest evidence will be accepted changes from the team's own task mix, measured with the corrections and oversight included.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model releases and technical results into clear stories developers can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).