What Is SWE-2 Flash? The Beginner's Guide to Cognition's Cheaper, Faster Coding Model

CodexFuse Team · 2026-11-05T11:29:58.425Z · 22 min read min read

What Is SWE-2 Flash? The Beginner's Guide to Cognition's Cheaper, Faster Coding Model

SWE-2 Flash is Cognition's smaller, cheaper, faster coding model for Devin, and this guide answers the question what is SWE-2 Flash in plain terms. Cognition launched it on November 4, 2026, one day before this guide was written, and it is available now inside Devin Desktop and the Devin CLI. The company says it runs at three times the speed and one-fifth the price of its larger sibling SWE-2, and it is free to try in Devin through November 30, 2026.

The launch matters because coding agents are only as useful as they are affordable to run. A model that is smart enough for routine work but cheap enough to run all day changes what teams automate, how often they ask an agent to review code, and how much documentation gets written by a machine instead of a person. This guide explains what SWE-2 Flash is, how Cognition built it, how its benchmarks compare with SWE-2 on both sides, what it costs, and where you can actually use it.

What Is SWE-2 Flash?

SWE-2 Flash is Cognition's smaller coding model for Devin, built by post-training Thinking Machines' Inkling-Small base model, and aimed at everyday engineering work rather than peak capability. It handles automations, code review, feature development, and documentation updates at a claimed three times the speed and one-fifth the price of SWE-2. It is the cheaper, faster sibling in Cognition's SWE-2 model family, and it is the first model in that family built on an American base model.

SWE-2 Flash explained in plain terms: if SWE-2 is the senior engineer you call for the hardest problems, SWE-2 Flash is the reliable junior who handles the steady stream of routine work quickly and cheaply. Cognition describes it as "a fully American coding model that sets a new bar for intelligence at low cost and high speed." The "fully American" label refers to its base model, Inkling-Small, which comes from Thinking Machines Lab, a United States lab, unlike SWE-2, which was built on Kimi K3 from Moonshot AI in China.

Flash is a mixture-of-experts model with 276 billion total parameters and 12 billion active parameters per token. Only about 4.3 percent of its parameters are active on any given token, since 12 divided by 276 is roughly 0.043, and that small active slice is a big part of why it runs fast and cheap. The model is Devin-only at launch: there is no standalone API, no public per-token price list, no model card, and no context window published for Flash as of November 5, 2026.

Before running any model file you download, check that it is the file the publisher actually released. A single changed byte can mean a tampered checkpoint, and the SHA-256 hash generator produces the hash you compare against the publisher's value. For a machine to run a model on, VPS hosting gives you a server you control, with the CPU and memory you choose.

How Was SWE-2 Flash Built?

SWE-2 Flash was built in two stages: a large-scale mid-training stage on top of Inkling-Small, followed by large-scale reinforcement learning. The mid-training stage is the new part of the recipe. Cognition says this is the first time it placed a large-scale mid-training stage before reinforcement learning, and the company credits that change, plus higher-quality reinforcement-learning tasks, for most of the gains over the base model.

Mid-training used a few trillion tokens of high-quality scientific articles, blogs, code, and other material, arranged in three stages with the scarcest, highest-quality data last. Cognition guided its data-mix choices with a curated held-out coding dataset and token-efficiency plots, and held-out validation loss improved at a stable log-linear rate through the stage. After mid-training, large-scale reinforcement learning took over, using audited instructions, environment setups, verifiers, and new task types.

The training also targeted cybersecurity work on purpose. Targeted security training raised the model's ExploitBench score from 11.3 percent for the base model to 34.8 percent for Flash, a gain of 23.5 points and roughly 3.1 times the starting score. Cognition says that dual-use capability is exactly why the weights are withheld: the same skills help defenders find flaws, but they could also help attackers, amid what the company describes as a recent surge in cyber attacks.

The result is a model that starts from a general-purpose base and ends as a specialist. Inkling-Small is a general multimodal model that takes text, image, and audio as input, while Flash is a coding model shaped by trillions of tokens of technical material and reinforcement learning on coding tasks. That specialization shows in the benchmark gains, which are large: Terminal-Bench 4.0 rose from 0.9 percent for the base model to 20.0 percent for Flash, about 22.2 times the starting score, and FrontierCode 2.0 rose from 3.2 percent to 24.0 percent, about 7.5 times.

Training a coding model is only half of a working agent system. The model also needs a way to call tools, read files, and hand work to other services, and that is the job of protocols such as the one described in what MCP over MOQT is, which moves model context over a fast transport. Flash itself runs inside Devin, where Cognition controls the whole agent loop.

SWE-2 Flash vs SWE-2: Which Model Wins What?

SWE-2 Flash vs SWE-2 is a cost-and-speed trade-off, not a simple upgrade. Flash is cheaper and faster, while SWE-2 is stronger on most benchmarks. Flash wins one of the five shared benchmark rows outright, near-ties a second, and trails on three, so the right choice depends on the job, not on which model is newer.

The two models differ in almost every dimension. Flash launched November 4, 2026, fifty-five days after SWE-2 launched on September 10, 2026. Flash is built on Inkling-Small, a 276-billion-parameter model, while SWE-2 is built on Kimi K3, a 2.8-trillion-parameter model, so SWE-2's base is about 10.1 times larger in total parameters and about 8.7 times larger in active parameters, since 104 billion active divided by 12 billion active is roughly 8.7.

The practical rule is simple. Send Flash the work that is common, well-specified, and high-volume: automations, code review, feature development, and documentation updates. Send SWE-2 the work that is hard, ambiguous, or mission-critical, where a few extra points of benchmark accuracy are worth paying five times as much and waiting longer.

A team that routes this way gets most of its agent work done at one-fifth the price and keeps the expensive model for the jobs that actually need it.

SWE-2 also has three reasoning-effort levels, Medium, High, and Max, trained in a single reinforcement-learning run. SWE-2 Medium takes 58 percent fewer turns and costs 81 percent less than SWE-1.7 on FrontierCode while scoring higher, with 53 mean steps per task against 127. Flash has no published reasoning-effort levels, which fits its role as the fast, routine-work model rather than the configurable flagship.

SWE-2 Flash Benchmarks: The Full Picture

SWE-2 Flash benchmarks are vendor-reported and mixed against SWE-2, so they need to be read in full, not cherry-picked. Cognition published the numbers on November 4, 2026, and no independent replication or hands-on review existed as of November 5, 2026. Every score below is the company's own run, and the honest summary is that Flash wins one comparison, near-ties another, and trails on three.

On FrontierCode 2.0, Flash scores 24.0 percent against SWE-2's 23.8 percent, an edge of 0.2 points. On SWE-Bench Pro v2 Hard, Flash scores 89.9 percent against SWE-2's 90.6 percent, a near-tie with a gap of 0.7 points. On the other three shared benchmarks Flash trails: FrontierCode 1.1 Main, 45.5 percent against 50.0 percent, a gap of 4.5 points; DeepSWE 1.1, 60.1 percent against 73.0 percent, a gap of 12.9 points; and Terminal-Bench 4.0, 20.0 percent against 27.3 percent, a gap of 7.3 points.

Two notes on sourcing. FrontierCode is Cognition's own benchmark, which asks whether a maintainer would merge the model's pull request, while Terminal-Bench and DeepSWE are independent agentic-coding benchmarks. The SWE-Bench Pro v2 Hard row comes from Cognition's table as reported by Superpower Daily, since the vendor post's extracted text did not expose its benchmark table directly.

SWE-Bench Pro v2 is Scale AI's harder successor to SWE-bench, with 642 tasks and a 51-task Hard split.

The same table also compares Flash with rival models, and that comparison cuts both ways. Flash beats the Chinese open-weight rivals only on the FrontierCode rows, while DeepSeek V4.1 Flash leads DeepSWE 1.1 at 77.0 percent and both Chinese models more than double Flash on Terminal-Bench 4.0, at 41.8 and 42.3 percent against 20.0 percent. Claude Haiku 5.5 edges Flash on both FrontierCode sets, at 27.7 and 46.4 percent.

Cognition's claim that Flash "outperforms all American open-weight models" is worded to exclude the Chinese models, so read it with that qualifier attached.

Against its own base model, the gains are large on every row. Terminal-Bench 4.0 moved from 0.9 percent for Inkling-Small to 20.0 percent for Flash. FrontierCode 2.0 moved from 3.2 percent to 24.0 percent.

DeepSWE 1.1 moved from 17.4 percent to 60.1 percent, about 3.45 times the starting score. ExploitBench moved from 11.3 percent to 34.8 percent. The mid-training stage plus reinforcement learning turned a general base model into a coding specialist, even if the result still sits below its much larger sibling.

How Fast Is SWE-2 Flash, Really?

SWE-2 Flash is three times faster than SWE-2 according to Cognition, but that figure describes generation speed, not end-to-end task time. The company's speed illustration replays each model's own run on the same Terminal-Bench 4.0 task, gsea-proteomics, at simulated rates of 200 tokens per second for Flash and 60 tokens per second for SWE-2, with transcripts shortened to equal length. The ratio of those simulated rates is 200 divided by 60, about 3.33, which is close to the claimed 3x.

The distinction matters. A simulated replay shows how quickly tokens stream out once the model is generating, but it does not measure how long each model takes to finish a whole task, including planning, tool calls, and retries. A model that generates tokens three times faster can still take similar wall-clock time on a task if it needs more turns or more attempts.

Treat "3x faster" as a statement about output speed inside Devin, not as a promise that every job finishes in one-third of the time.

Even with that caveat, speed is the point of Flash. Coding agents spend much of their time waiting on the model, and a faster model means shorter feedback loops for the developer watching the agent work. For routine work such as documentation updates or small refactors, where the task is short and the model rarely gets stuck, faster generation translates directly into a snappier experience.

For hard multi-step tasks, where the model may need many turns, the speed advantage matters less than accuracy, which is why SWE-2 keeps its role.

SWE-2 Flash Devin: Where You Can Use It

SWE-2 Flash in Devin is available in Devin Desktop and the Devin CLI only, from November 4, 2026. Devin is Cognition's AI software engineer, an autonomous agent that plans, writes, runs, and tests code and opens pull requests, introduced on March 12, 2024 as "the first AI software engineer." Devin Desktop is the agentic IDE, formerly Windsurf, which Cognition acquired in July 2025 and renamed Devin Desktop on June 2, 2026. The Devin CLI brings Devin to the terminal, working on local files, with a /handoff command to pass longer tasks to Devin Cloud.

Flash is not announced for Devin Cloud, Devin Web, or Fusion at launch, so web and cloud users cannot pick it yet. SWE-2's own Web and Fusion rollout was described as rolling out at its launch, and there is no statement that Flash follows the same path or timeline. There is also no standalone API and no API identifier published, so Flash cannot be called from outside Devin at all.

To try Flash in Devin, follow these five steps:

1. Open Devin Desktop or the Devin CLI on a machine where Devin is already set up.
2. Pick SWE-2 Flash as the model for your session, where the model picker lists it.
3. Give the agent a routine job first, such as a documentation update or a code review, to see how it handles everyday work.
4. Check the result the same way you would check any agent's work, by reading the diff and running the tests.
5. Keep using Flash while the free window lasts, and switch to SWE-2 for any job that needs the stronger model.

Devin is one of several terminal and IDE coding agents, and the skills transfer. If you want a grounding in the most widely used one first, read how to use Claude Code, which covers the same loop of prompting, reviewing diffs, and running tests that Flash follows inside Devin.

SWE-2 Flash Price: What It Costs and the Free Window

SWE-2 Flash price is simple to state and easy to overstate: Cognition says it costs one-fifth as much as SWE-2 inside Devin, and it is free in Devin Desktop and the Devin CLI through November 30, 2026. That free window is a time-limited offer, not a permanent price, and it ends 25 days after this guide was written on November 5. Coverage of the launch specifies that the free offer applies to paid Devin users on Pro, Max, and Teams plans, so do not read it as free for everyone.

No per-token price has been published for Flash. Cognition has published no dollar figure per million tokens, no API price list, and no price that applies after November 30. The only price statement is relative: one-fifth of SWE-2 inside Devin.

Anyone quoting a Flash dollar price is inventing one.

For context only, SWE-2's Devin list price is $3.00 input, $15.00 output, and $0.30 cached input per one million tokens, according to Devin documentation reproduced by secondary sources. One-fifth of those numbers would be $0.60 input and $3.00 output, but that arithmetic is ours, not Cognition's, and it must not be quoted as Flash's published price. SWE-2 itself is also free on self-serve Desktop and CLI through November 30, 2026, then 75 percent off list, or $0.75 input and $3.75 output, through December 31, 2026.

Devin plan pricing reported in the same coverage is $20 per month for Pro and $200 per month for Max.

The honest way to think about the price is per job, not per token. If a team runs hundreds of routine agent jobs a week, paying one-fifth per job changes which jobs are worth automating. A documentation update that was too expensive to hand to an agent at SWE-2 prices becomes cheap at Flash prices, while a hard debugging session still justifies the full SWE-2 rate.

The free window through November 30 is the right time to measure that difference on your own workload, before any price takes effect.

Is SWE-2 Flash Open Source?

SWE-2 Flash is not open source, and its trained weights are not publicly downloadable. Cognition says it is building a program to securely share the weights with trusted industry partners and research labs, with more details promised in a few weeks, but the public cannot download Flash today. The reason given is cybersecurity: the same targeted training that raised ExploitBench from 11.3 percent to 34.8 percent could help attackers as well as defenders.

The confusion comes from the base model. Inkling-Small, the model Flash was built on, is open: Thinking Machines released it on July 30, 2026, ninety-seven days before Flash launched, under Apache 2.0, with downloadable weights in BF16 and NVFP4 formats on Hugging Face, and a 1-million-token context window. That openness applies to the base model only.

Flash is Cognition's mid-trained and reinforcement-learning post-trained derivative, and its weights stay private, so never present Inkling-Small's open weights as Flash being open.

The gated sharing program is Cognition's answer to the tension between transparency and safety. Partners and research labs get access under a controlled program, while the public does not, and Cognition keeps the distribution switch. Whether that model satisfies researchers who want to reproduce results is an open question, and no independent replication of Flash's benchmarks existed as of November 5, 2026.

What Is SWE-2, the Parent Model?

SWE-2 is Cognition's most advanced coding model, launched on September 10, 2026, and it is the larger sibling Flash is measured against. Cognition post-trained it from Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight mixture-of-experts model with 104 billion active parameters, which had already been through extensive reinforcement learning for agentic coding. Cognition says its own reinforcement learning adds five to six points on many benchmarks.

SWE-2's headline result is 50.0 percent on FrontierCode 1.1 Main, within one point of Fable 5.1 at 50.9 percent while being 64 percent cheaper, according to Cognition's launch framing. Its own launch table also shows 73.0 percent on DeepSWE 1.1, 92.8 percent on Terminal-Bench 2.1, and 27.3 percent on Terminal-Bench 4. SWE-2 is available in Devin Desktop and CLI, with Devin Web and Fusion rolling out, and like Flash it has no standalone API and no open weights.

The family is now two models with distinct jobs. SWE-2 is the flagship for hard work, and Flash is the workhorse for routine work. Cognition's earlier model, SWE-1.7, sits below both on the launch table, at 42.0 percent on FrontierCode 1.1 Main and 7.6 percent on Terminal-Bench 4.

The progression from SWE-1.7 to SWE-2 to SWE-2 Flash shows a company that first pushed capability, then pushed cost and speed.

SWE-2 Flash vs Other "Flash" Models

SWE-2 Flash vs other "Flash" models is a naming collision, not a rivalry. "Flash" is also used by Google's Gemini Flash models and by DeepSeek V4.1 Flash, and those are different companies' models with different designs. In this guide, "Flash" always means Cognition's SWE-2 Flash unless stated otherwise.

The collision matters because the benchmark table itself includes two of the namesakes. DeepSeek V4.1 Flash scores 77.0 percent on DeepSWE 1.1 and 42.3 percent on Terminal-Bench 4.0, ahead of SWE-2 Flash on both. GLM 5.3 Flash scores 41.8 percent on Terminal-Bench 4.0, more than double SWE-2 Flash's 20.0 percent.

A reader who searches for "Flash benchmarks" without the SWE-2 prefix can easily land on the wrong model's numbers, so always check the full model name before quoting a score.

The same care applies to SWE-bench. SWE-2 is a coding model, not a benchmark, and SWE-2 Flash is a smaller sibling model, not a version of SWE-bench. SWE-bench and SWE-bench Verified are the original academic benchmarks that grade models on resolving GitHub issues, while SWE-Bench Pro and SWE-Bench Pro v2 are Scale AI's harder successors.

Flash's table does not report SWE-bench Verified at all, and its 89.9 percent figure is on the SWE-Bench Pro v2 Hard split only.

Who Should Use SWE-2 Flash, and Who Should Stick With SWE-2?

SWE-2 Flash suits developers and teams whose agent work is routine, well-specified, and high-volume. If your week is full of documentation updates, code reviews, small feature additions, and automations that follow a clear pattern, Flash gives you most of the capability at one-fifth the price and three times the speed. The free window through November 30 makes it cheap to test on your own codebase before committing.

SWE-2 still suits the hard jobs. If a task is ambiguous, spans many files, needs careful judgment, or is mission-critical, the larger model's benchmark lead is worth paying for: 12.9 points on DeepSWE 1.1, 7.3 points on Terminal-Bench 4.0, and 4.5 points on FrontierCode 1.1 Main. A wrong answer on a hard task costs more than the price difference, because a developer has to find the mistake, undo the change, and redo the work.

Small teams and solo developers gain the most from Flash, since the price cut makes agent use affordable at volume. Larger teams gain a routing decision: send routine work to Flash by default and escalate to SWE-2 when the task is hard. Security-minded teams should note the other side of the trade, which is that Flash's weights are private and its cybersecurity training is dual-use, so the model itself is a closed product even though its base model is open.

Running agents in a safe place matters as much as picking the model. A sandbox keeps an agent's experiments away from your real data, and how to use E2B shows how to give an agent a disposable cloud sandbox to work in. Flash inside Devin plus a sandbox for execution is a sensible pairing for teams that want speed without giving an agent the keys to production.

The Bottom Line: Should You Try SWE-2 Flash?

SWE-2 Flash is worth trying if you already use Devin, because the cost of trying is zero until November 30, 2026. Open Devin Desktop or the Devin CLI, pick Flash, and give it a routine job you would normally do by hand. Measure the result the way you would measure any tool: did it finish the job, was the diff clean, and did the tests pass?

If the answer is yes on routine work, Flash can carry most of your agent load at one-fifth the price.

Keep expectations honest. Flash is not a better SWE-2, and Cognition's own numbers say so: it trails on three of five shared benchmarks and only edges SWE-2 on one. What it is, is a cheaper, faster model for the large share of engineering work that does not need the flagship's full strength.

That is a useful product, and it is the first Cognition model built on an American base, with a new mid-training recipe behind it.

The clock is the one hard fact. The free window ends November 30, 2026, twenty-five days after this guide was written, and no Flash price has been published for after that date. Try it while it is free, learn which of your jobs it handles well, and decide with your own evidence before the price appears.

Sources

Cognition — Introducing SWE-2 Flash (November 4, 2026)

Unite.AI — Cognition Launches SWE-2 Flash for Faster, Lower-Cost Coding Agents (November 4, 2026)

RuntimeWire — Cognition launches SWE-2 Flash in Devin at one-fifth the price of SWE-2 (November 4, 2026)

AI Parabellum — Cognition Launches SWE-2 Flash, a Cheaper, Faster Coding Model for Devin (November 4, 2026)

Superpower Daily — Cognition adds a lower-priced coding model to Devin, keeping its weights private (November 5, 2026)

StartupFortune — Cognition launches SWE-2 Flash to make AI coding agents cheap and fast (November 5, 2026)

CellCog — Cognition SWE-2 and SWE-2 Flash: Benchmarks and Cost (November 2026)

Thinking Machines — Inkling-Small model card (July 30, 2026)

Frequently Asked Questions

What is SWE-2 Flash?

SWE-2 Flash is Cognition's smaller, cheaper, faster coding model for Devin, launched November 4, 2026. It is built by post-training Thinking Machines' Inkling-Small base model and aimed at everyday engineering work, including automations, code review, feature development, and documentation updates, at three times the speed and one-fifth the price of SWE-2.

Is SWE-2 Flash free?

Yes, temporarily. SWE-2 Flash is free in Devin Desktop and the Devin CLI through November 30, 2026, according to Cognition's launch post. Coverage of the launch specifies that the offer applies to paid Devin users on Pro, Max, and Teams plans.

After that date no Flash price has been published, only that it costs one-fifth of SWE-2 inside Devin.

Is SWE-2 Flash better than SWE-2?

No, not across the board. SWE-2 Flash is a cost-and-speed trade-off: on Cognition's own vendor-reported numbers it narrowly wins FrontierCode 2.0, 24.0 percent against 23.8 percent, and near-ties SWE-Bench Pro v2 Hard, 89.9 percent against 90.6 percent, but it trails on Terminal-Bench 4.0, DeepSWE 1.1, and FrontierCode 1.1 Main. Use Flash for routine work and SWE-2 for hard tasks.

How much does SWE-2 Flash cost?

No per-token price has been published for SWE-2 Flash. Cognition only says it costs one-fifth as much as SWE-2 inside Devin, and it is free in Devin Desktop and the Devin CLI through November 30, 2026. Do not quote a dollar figure for Flash; SWE-2's Devin list price of $3 input and $15 output per one million tokens is a different model's price.

Can I download SWE-2 Flash, or is it open source?

No. The trained SWE-2 Flash weights are not publicly downloadable, and Cognition is building a gated weight-sharing program for trusted industry partners and research labs instead. Only the base model, Inkling-Small, is open under Apache 2.0 and downloadable from Hugging Face, and that is not Flash.