Cognition says its new SWE-2 Flash model brings faster, lower-cost coding-agent work to Devin Desktop and CLI. The company reports a threefold speed advantage and one-fifth the price of its SWE-2 model, but its own benchmark table shows that faster and cheaper does not mean stronger on every software task. The model is available inside Devin, while public access to its weights is not part of the launch.
What Cognition launched
On November 4, Cognition introduced SWE-2 Flash, a coding model offered through its Devin Desktop application and command-line interface. The release positions Flash as a smaller, more economical companion to SWE-2 for routine software-agent tasks. Cognition names automation, code review, feature development and documentation updates as examples of work it expects smaller models to handle.
The announcement is notable less as a claim of universal benchmark leadership than as a product trade-off: Cognition says a model can serve a broad share of practical agent work while using less time and cost per interaction. Those are the company’s descriptions of its product and its intended use, not an independent finding that the model is the best choice for a given engineering team. The official announcement says Flash is available now in Devin Desktop and CLI and is free through November 30, 2026.
Cognition describes SWE-2 Flash as “fully American” and says it is based on Inkling-Small, an open-weight model from Thinking Machines. According to the company, Flash has 276 billion total parameters and 12 billion active parameters. That parameter configuration is also reported for Inkling-Small. The announcement says Cognition added a mid-training stage and then reinforcement learning; it does not publish model weights or a complete independent reproduction recipe with which outsiders could recreate the result.
Benchmarks: a mixed comparison
The scores Cognition published do not show a clean sweep over its larger SWE-2 model. On FrontierCode 2.0, SWE-2 Flash scores 24.0, narrowly ahead of SWE-2 at 23.8. On Terminal-Bench 4.0, Flash scores 20.0 while SWE-2 scores 27.3. On DeepSWE 1.1, Flash records 60.1 against SWE-2’s 73.0. The pattern makes the company’s positioning as a speed-and-price alternative more defensible than a claim that Flash replaces its larger sibling across coding evaluations.
Against its reported base model, Inkling-Small, Flash posts higher figures on all three of those comparisons: the base has 3.2 on FrontierCode 2.0, 0.9 on Terminal-Bench 4.0 and 17.4 on DeepSWE 1.1. Cognition attributes improvement to its training process. These are values in the company’s table, however; neither the launch announcement nor the coverage supplied here offers an independent replication or a detailed third-party audit of benchmark runs.
Across the wider table, the trade-offs persist. Flash is listed at 45.5 on FrontierCode 1.1 Main, below SWE-2’s 50.0; at 89.2 on Terminal-Bench 2.1, below SWE-2’s 92.8; and at 89.9 on SWE-Bench Pro v2 Hard, just below SWE-2’s 90.6. The table also gives Flash 50.0 on SWE-Atlas QA versus SWE-2’s 56.3. On ExploitBench, Flash scores 34.8 versus 36.1 for SWE-2. Benchmarks measure defined task sets and should not be treated as direct forecasts of an individual team’s results.
Cognition says Flash outperforms all American open-weight models on the leading evaluations it reports, including models three times larger. That conclusion depends on the company’s chosen model set and evaluations. The announcement does not provide independent verification of the comparative claim. The Unite.AI report reproduces the principal numbers and describes their trade-offs, but identifies the scores as Cognition-reported data rather than presenting fresh testing.
Speed, price and intended work
Cognition advertises Flash at three times the speed and one-fifth the price of SWE-2. Its speed illustration compares generation rates of 200 tokens per second for Flash and 60 for SWE-2. The company says it replays each model’s own run on Terminal-Bench 4.0’s gsea-proteomics task at those simulated generation rates, shortening transcripts to equal length. That illustration is not the same as a broad, independently measured end-to-end latency study across varied codebases and agent workflows.
The distinction matters for software agents. Token generation rate is one component of elapsed task time; total time can also depend on tool calls, test execution, retries, context size and the agent’s ability to reach a correct result. A faster stream of generated tokens does not by itself establish faster completion of every task. Similarly, the “one-fifth the price” claim is Cognition’s stated comparison. The announcement does not give a full tariff breakdown in the text presented, so prospective users should consult Devin’s current pricing and their own usage terms before treating the ratio as a universal budget estimate.
For users already working in Devin, the immediate question is whether Flash performs adequately on frequent, lower-stakes tasks. Code review, small changes, documentation and repetitive automation may be candidates for trial, but the company’s examples are use cases, not independent guarantees. A practical evaluation would compare the same representative tasks across models, record completion quality and human correction time, and include tool and compute costs. Teams handling high-impact changes may prefer a stronger result on their own task set over a speed headline.
How Cognition says it trained Flash
Cognition says the model starts from Inkling-Small, then adds large-scale mid-training before reinforcement learning. It reports that this mid-training used a few trillion tokens of scientific articles, blogs, code and other sources. The team says it curated a held-out coding dataset meant to reflect real software-engineering scenarios and compared data mixtures by how many tokens they needed to reach a baseline validation loss. It says training occurred in three stages, with the scarcest and highest-quality data introduced later.
After mid-training, Cognition says it used large-scale reinforcement learning and that most gains came from greater quality and diversity in its tasks, rather than major changes to the reinforcement-learning recipe it had used for SWE-2. It says it audited task instructions, environment setup and verifiers and added types of work meant to broaden reasoning and problem-solving. These are descriptions of the company’s internal process. The announcement gives a high-level account but not the complete datasets, task suite, model checkpoints or independent analysis needed to assess each contribution separately.
The company also says targeted cybersecurity training raised Flash’s ExploitBench score from Inkling-Small’s 11.3% to 34.8%. It frames the result as evidence of increased cybersecurity capability and also as a reason to be cautious about releasing weights. That score is an evaluation result reported by Cognition, not proof that the model is either safe or unsafe in deployment. Benchmark performance alone cannot characterize how the model will behave across real security contexts.
Weights, security and access plans
SWE-2 Flash is not being released publicly as downloadable weights. Cognition says it is building a secure-access program for trusted industry partners and research labs, with more details to come in a few weeks from the announcement. The company’s stated aim is to let approved users fine-tune and run the model in sensitive environments, including on-premises infrastructure, while managing risks associated with cyber capabilities.
This is a plan, not a currently documented access offer. The launch post does not specify eligibility rules, review procedures, technical safeguards, audit rights, availability dates or what partners will be allowed to do. Until Cognition publishes those terms, organizations should not assume they can obtain or host Flash weights. Nor should “trusted partner” be read as a defined category with publicly stated criteria.
The policy reflects a genuine tension that the announcement itself recognizes: more access can support evaluation, customization and deployment control, while the same capabilities may be misused. Cognition points to its higher ExploitBench result and refers to cyber risks as reasons not to make the weights public. Whether a gated program manages that balance will depend on its eventual controls and accountability, details that are still pending.
What the announcement does—and does not—establish
The release establishes that Cognition has announced Flash within Devin Desktop and CLI, and it documents the company’s own rationale, score table and access intentions. It also gives readers a useful comparison with SWE-2: Flash has one small lead on FrontierCode 2.0 among the three headline evaluations, but trails on Terminal-Bench 4.0 and DeepSWE 1.1. This specificity is more informative than an undifferentiated claim that a model is “better.”
It does not establish, independently, that the advertised price and speed ratios hold for all workloads, that Flash outperforms every relevant American open-weight model, or that its benchmark scores predict a team’s productivity. The supplied independent article from Unite.AI summarizes the announcement and reproduces its core comparison, but it is not an independent benchmark report. Its coverage is useful confirmation that the launch was reported on November 4, not a second source validating the performance figures.
There is no apparent topic-level duplicate in the publication’s recent post list. Recent coverage includes a study about coding-agent harness costs and an OpenAI model launch, both adjacent to agent economics and model releases, but neither covers Cognition’s SWE-2 Flash announcement. The candidate is therefore distinct enough to publish if the newsroom treats the metrics as attributed company claims and makes the planned nature of secure weight access explicit.
Frequently asked questions
Where can people use SWE-2 Flash?
Cognition says it is available in Devin Desktop and Devin CLI. Its announcement says access is free through November 30, 2026; users should check Devin for current terms.
Is it faster and cheaper than SWE-2?
Cognition reports three times the speed and one-fifth the price. These are company-reported comparisons, and the supplied sources do not independently benchmark or verify them across workloads.
Does SWE-2 Flash beat SWE-2 on benchmarks?
Not across the headline scores. Cognition reports 24.0 versus 23.8 on FrontierCode 2.0 in Flash’s favor, but 20.0 versus 27.3 on Terminal-Bench 4.0 and 60.1 versus 73.0 on DeepSWE 1.1.
Can researchers download the model weights?
No public weights are offered in the announcement. Cognition says a secure program for trusted industry partners and research labs is being developed, with further details planned.
Are the benchmark claims independently confirmed?
Not by the two supplied sources. Cognition published the results, and Unite.AI repeats the company’s table; neither source provides a separate benchmark replication.