Cognition's SWE-2 Flash Runs 3x Faster at One-Fifth the Cost

Cognition has released SWE-2 Flash, a smaller coding model for agents that edit repositories and operate development tools in Devin. The company built it from Thinking Machines’ Apache 2.0-licensed Inkling-Small, added a multi-stage mid-training pass over several trillion tokens, and finished with reinforcement learning on verified agent tasks. Cognition reports substantial benchmark gains over the base model, although public API access and unrestricted weights downloads remain unavailable.

276B parameters, 12B active

SWE-2 Flash retains Inkling-Small’s sparse Mixture-of-Experts architecture. A Mixture-of-Experts model contains many specialized parameter groups but activates only a subset for each token, reducing computation while preserving a larger pool of learned capacity.

Each token in Flash routes to six of 256 specialist experts, alongside two shared experts that process every token. Activating about 12 billion parameters lowers the arithmetic required for generation. Serving costs still include memory bandwidth, expert routing, attention, and key-value cache overhead, so its performance will differ from that of a conventional dense 12B model.

Cognition reports large improvements after mid-training and reinforcement learning:

The speed claim: 3x inside Devin

In Cognition’s demonstration of a Terminal-Bench 4.0 task, SWE-2 Flash generated about 200 tokens per second, compared with 60 for SWE-2. The company summarizes the product as three times faster and one-fifth the in-product price of its larger model. End-to-end agent latency also depends on prompt processing, tool calls, command execution, and repository size, so a single generation demo does not establish performance across workloads.

Access is limited to Devin Desktop and the Devin CLI. Cognition’s launch post says Pro, Max, and Teams subscribers can use Flash without an additional charge through November 30. The company has not announced a public API or a timetable for broader availability.

The trained weights remain gated

The downloadable component is Inkling-Small, the Apache 2.0 base model. Cognition plans to provide the trained SWE-2 Flash checkpoint only to selected partners and researchers, preventing independent developers from inspecting, fine-tuning, or self-hosting the released model. Apache 2.0 permits proprietary derivatives, so the base model’s license does not require Cognition to publish its resulting checkpoint.

Cognition cites cybersecurity risk as the reason for selective access. Its training included security tasks that raised the model’s ExploitBench result from 11.3% to 34.8%. The company intends the partner program to support controlled research and on-premises deployment while limiting distribution of the enhanced checkpoint.

Cognition describes Flash as leading U.S.-developed models in its open-weight comparison group. That label refers to the peer set used for evaluation; unrestricted access to the trained SWE-2 Flash weights is unavailable.

A new layer between pretraining and RL

Cognition inserted a mid-training phase between Inkling-Small’s original pretraining and task-level reinforcement learning. Mid-training continues next-token learning on a more focused corpus, allowing the model to absorb domain material before reinforcement learning rewards successful behavior on complete tasks.

1. Start with Inkling-Small as the pretrained base.
2. Continue training on several trillion tokens of code, scientific articles, and technical blogs.
3. Split that corpus into three stages, reserving the scarcest, highest-quality material for the last stage.
4. Run large-scale reinforcement learning on curated agent tasks with audited environments and verifiers.

Cognition selected its data mixture with token-efficiency experiments. For each candidate mixture, the team measured how many training tokens were required to reach a fixed validation loss, then compared that figure with a baseline. A mixture that reached the target with fewer tokens delivered more learning for the same training budget.

The reinforcement-learning work concentrated on task quality and diversity. Researchers audited instructions, environment setup, and verifiers, which are automated checks that decide whether an agent completed a task correctly. Weak verifiers can reward shortcuts or reject valid solutions, so their design directly affects what the model learns. Cognition also added tasks covering general reasoning and problem-solving beyond repository patching.

Strong patches, weaker shell work

Cognition’s evaluation table shows the strongest results on software-engineering tasks aligned with its reinforcement-learning curriculum. The figures remain company-reported, and independent replication is constrained by the absence of a public API or checkpoint.

The Terminal-Bench 2.1 and 4.0 scores come from different benchmark versions with different task sets, making direct comparison inappropriate. Within Terminal-Bench 4.0, Flash’s gap behind GLM and DeepSeek indicates weaker performance on that evaluation’s multi-step shell tasks. Its SWE-Atlas QA result also points to lower code-retrieval and question-answering performance than its repository-editing scores suggest.

A volume model for Devin

Cognition positions SWE-2 Flash for routine agent workloads such as automations, code review, feature development, and documentation updates. The larger SWE-2 remains the option for more demanding tasks, while Flash offers lower generation cost and latency for higher-volume work. Existing Devin subscribers can evaluate that tradeoff during the promotional period, but teams outside Devin currently have no supported integration path.

Inkling-Small also shows why the base-model choice matters. It uses 276 billion total parameters and 12 billion active parameters, under one-third of the 975 billion total and 41 billion active parameters in the larger Inkling model. Cognition says the smaller model occupies a similar tier on the cited Intelligence Index and leads open-weight models at its size or below.

The development strategy shifts investment from general-purpose pretraining toward domain-focused data, verifier design, and reinforcement learning. SWE-2 Flash provides Cognition’s evidence that a strong open base can become a specialized coding agent with lower inference requirements, even though the final checkpoint remains a gated product.