Multiple reasoning-effort levels
SWE-2 supports medium, high, and max effort levels, allowing the available reasoning effort to be matched to task complexity and cost considerations.
SWE-2
SWE-2 is Cognition's agentic coding model for planning, implementing, testing, and verifying software changes. It supports multiple effort levels and is available through Devin Desktop and CLI, with rollout stated for Devin Web and Fusion.
SWE-2 overview
SWE-2 is Cognition's coding model for agentic software-engineering work. It is designed to plan and execute code changes while balancing coding performance, inference cost, and rollout time across multiple reasoning-effort levels.
The model is available in Devin Desktop and CLI, with the source article also stating that rollout is underway for Devin Web and Fusion. Cognition presents SWE-2 as a model for software workflows where an agent needs to inspect a codebase, implement changes, write tests, and verify the resulting work.
Key features
Focused codebase exploration
The model focuses exploration on the parts of a codebase that matter for the task, helping it begin implementation sooner than SWE-1.7 in the cited evaluation.
Broader test coverage
SWE-2 is reported to write more end-to-end tests and check regressions and edge cases more reliably during coding tasks.
Resourceful task execution
When an expected route is blocked, the model can seek another path within the user's stated boundaries, including reconstructing data from accessible Slack history in one reported example.
Verification-oriented behavior
The model re-derives conclusions, verifies hypotheses, and runs artifacts to gather evidence rather than relying only on surface-level descriptions.
Cost-aware post-training
Its post-training uses cost-penalized rewards tuned to the base model's cost-performance curve, with the goal of improving performance across effort levels in one RL run.
Potential use cases
Routine code changes
Use the medium effort level for straightforward fixes or intermediate coding tasks where quick progress and lower inference cost are priorities.
Complex implementation work
Assign high or max effort to larger or more uncertain engineering tasks that benefit from additional planning, codebase exploration, and verification.
Test and regression work
Have the agent create end-to-end tests and check regressions or edge cases as part of an implementation workflow.
Blocked investigations
Use the model to investigate a blocked task by looking for an alternative route within the data and tools the user has authorized.
Evidence-based debugging
Ask the agent to validate a hypothesis by re-deriving the conclusion and running relevant artifacts rather than accepting an unverified explanation.
Pros and Cons
Pros
- Offers several effort levels for different cost and task-complexity tradeoffs.
- Reportedly reduces exploratory turns and average cost relative to SWE-1.7 on the cited FrontierCode 1.1 Main evaluation.
- Adds behaviors focused on end-to-end testing, alternative paths, and evidence-based verification.
- Available across Devin Desktop and CLI, with stated rollout to Devin Web and Fusion.
Cons
- The available source does not provide pricing, plan limits, integrations, setup requirements, security details, or team-administration information.
- Benchmark results vary by task and effort level, so the headline scores do not establish performance for every software project.
FAQ
Where is SWE-2 available?
SWE-2 is available in Devin Desktop and CLI, with rollout to Devin Web and Fusion stated in the source article.
What is SWE-2?
SWE-2 is a coding model post-trained from Kimi K3 and designed for agentic software-engineering tasks across multiple reasoning-effort levels.
How do SWE-2 effort levels differ?
The source describes medium, high, and max effort levels. Medium is positioned for more cost-efficient performance on simple and intermediate tasks, while high and max spend more effort on complex tasks.
How does SWE-2 compare with SWE-1.7 on efficiency?
On FrontierCode 1.1 Main, SWE-2 medium scored higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average, according to the source article.
Quick Facts
- Category
- Developer tool
- Product type
- Agentic coding model
- Platforms
- Devin Desktop and CLI; rollout stated for Devin Web and Fusion
- Effort levels
- Medium, high, and max
- Base model
- Kimi K3, a 2.8T-parameter model
- FrontierCode 1.1 Main score
- 50.0% for SWE-2 in the cited benchmark
SWE-2 대안
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repositories and verifies changes with the real compiler, debugger, sanitizers, and tests before showing a diff. It offers a free tier plus paid plans, with editor connectors and zero-retention handling described in the source.
Ghost
Ghost는 채팅, 코드 생성, 터미널 작업 실행을 지원하는 터미널 기반 AI assistant입니다. 무료 모델을 제공하며 Linux, macOS, Windows를 지원하는 오픈 소스입니다.
Vi3ecode
Vi3ecode is a maintained development environment for working with AI coding agents. It keeps project context, terminal, Git, memory, flows, review, and communication connected around the agent you already use.
Lucid Train
Lucid Train is a local-first AI coding harness that turns repository code into architecture diagrams and uses those diagrams as specifications for coding agents. It is available as a desktop app and a lightweight Rust CLI, and it can run with local models or your own API keys.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads inside Firecracker micro-VMs. It is designed for workflows that need machine-level isolation, private networking between sandboxes, and programmatic control through SDK, CLI, or MCP.
hob
hob is an independent workspace for coding agents that keeps agent sessions, terminals, history, and follow-up work organized around the tools and providers you already use. It is aimed at developers who want local control over routing, history, and workspace structure rather than a bundled model stack.