SWE-2 Flash: Cognition's cheaper coding model for Devin, claims checked

Cognition, the company behind Devin, released SWE-2 Flash on November 4, 2026: a smaller coding model it sells at one-fifth the price and three times the speed of its flagship SWE-2. It is available now inside Devin Desktop and the Devin CLI, and nowhere else. The interesting part is not the marketing line about "intelligence at low cost." It is the training recipe (a mid-training stage before reinforcement learning, which Cognition says it had not done before) and the honest-looking benchmark table, which shows a model that wins on some agentic evals and clearly loses on others. If you run coding agents in production, this is worth understanding because routine agent work (code review, doc updates, small features) is exactly the workload Cognition is pricing down.

What Cognition actually shipped

SWE-2 Flash is a mixture-of-experts model with 276B total parameters and 12B active parameters, built on top of Inkling-Small, a base model from Thinking Machines. Cognition calls it a "fully American coding model," which is vendor positioning, not a technical property.

The headline claims, all from Cognition's announcement post:

• One-fifth the price and 3x the speed of SWE-2, its much larger sibling (listed in the table at 2.8T total, 104B active parameters).
• A replay comparison in the post simulates SWE-2 Flash at 200 tokens per second against SWE-2 at 60 tokens per second. Treat this as an illustration of the speed gap, not a measured guarantee for your workload.
• Target workloads: the routine tasks coding agents handle constantly, including code review, workflow automation, feature development, and documentation updates.
• Availability: Devin Desktop and the Devin CLI, starting November 4, 2026. The Devin Desktop page confirms the product surface exists, with a macOS download and local plus cloud agent management.

The weights are not public. More on that below, because it is the most consequential line in the post.

The benchmark table, read carefully

Cognition published a comparison table against its own base model (Inkling-Small), its flagship (SWE-2), and a row of external models including GLM 5.3 Flash, DeepSeek V4.1 Flash, GPT-6 Luna, and Claude Haiku 5.5. I cannot verify the external models' numbers; Cognition ran these evals itself. What the table shows, per Cognition:

• FrontierCode 2.0: SWE-2 Flash scores 24.0, slightly ahead of SWE-2 at 23.8 and GPT-6 Luna at 21.9, behind Claude Haiku 5.5 at 27.7. The mid-training stage alone moved the base model from 3.2 to 24.0 on this eval.
• Terminal-Bench 4.0: SWE-2 Flash scores 20.0, behind SWE-2 at 27.3 and well behind Claude Haiku 5.5 at 39.2, GLM 5.3 Flash at 41.8, and DeepSeek V4.1 Flash at 42.3. The base model scored 0.9, so mid-training did heavy lifting here, but the model still trails the field.
• DeepSWE 1.1: SWE-2 Flash scores 60.1 against SWE-2's 73.0 and DeepSeek V4.1 Flash's 77.0.
• Terminal-Bench 2.1: 89.2, close to SWE-2's 92.8.
• SWE-Bench Pro v2 Hard: 89.9, nearly tied with SWE-2 at 90.6.
• ExploitBench, a cybersecurity eval: 34.8, up from the base model's 11.3. Cognition says it deliberately targeted cyber capability in training to help defenders.

Read generously, SWE-2 Flash roughly matches its much larger sibling on SWE-Bench Pro v2 Hard and Terminal-Bench 2.1 while running at a fraction of the cost. Read skeptically, it loses to SWE-2 on DeepSWE 1.1 and Terminal-Bench 4.0 by wide margins, and it is not close to the frontier models Cognition itself lists on several evals. Both readings come from the same vendor table. The claim that it "outperforms all American open-weight models, even models 3x larger" is Cognition's framing of its own runs.

The practical takeaway: for the routine, high-volume tasks where per-call cost dominates, a model that is nearly tied with the flagship on SWE-Bench Pro v2 Hard at one-fifth the price is a real option. For hard terminal-heavy tasks, the vendor's own numbers say to keep the bigger model.

How they trained it

This is the part worth stealing for your own eval thinking. Cognition added a mid-training stage between the base model and reinforcement learning, which it says is new to its recipe.

The process, per the post:

1. They curated a held-out coding dataset meant to reflect real software engineering scenarios, used to judge data quality.

2. They compared candidate data mixes with token efficiency plots: for each mix, how many tokens it takes to reach the same validation loss as a baseline mix. Mixes that need fewer tokens win.

3. The converged dataset was a few trillion tokens of scientific articles, blogs, code, and other sources, trained in three stages so the scarcest, highest-quality data came last. Validation loss improved at a stable log-linear rate through each stage.

4. Then large-scale RL, with most of the gains attributed to higher quality and diversity of RL tasks rather than recipe changes: audited instructions, environment setup, and verifiers, plus new task types.

Two things stand out. First, mid-training on a 12B-active MoE moved Terminal-Bench 4.0 from 0.9 to 20.0, which suggests the base model was badly undertrained for agentic terminal work rather than incapable of it. Second, the token-efficiency methodology is a cheap, reproducible way to argue about data mixes with evidence instead of taste.

The weights question

Cognition says it wants to give industry partners and researchers transparent access to weights for sensitive and on-prem deployments, and that it raised the model's ExploitBench score from 11.3 to 34.8 specifically to help cyber defenders. It also says those same capabilities can help attackers, citing a recent surge in cyber attacks, so it is not releasing weights publicly. Instead, it is building a program to share weights securely with trusted partners and labs, with more detail promised in the coming weeks.

Until that program exists, SWE-2 Flash is a managed-only model inside Devin. If your constraint is self-hosting, this release changes nothing for you today.

What I would check

These are my untested suggestions, not findings:

• Run your own 50 to 100 representative routine tasks (a code review set, a doc update set, a small feature set) through both SWE-2 Flash and whatever you use today, and compare cost per accepted output, not per token. Vendor speed claims rarely survive contact with real retry loops.
• Check how Devin bills Flash versus SWE-2 on your plan, and whether "one-fifth the price" refers to list price per token or to observed cost per completed task.
• If you are in a regulated environment, track the weights-sharing program announcement. On-prem availability would change the compliance math.
• Watch whether the FrontierCode 2.0 edge over SWE-2 (24.0 vs 23.8) survives independent replication. A 0.2 point gap is within noise on most eval harnesses.

What I would not conclude

• That SWE-2 Flash beats SWE-2. The vendor table shows it losing on DeepSWE 1.1 and Terminal-Bench 4.0 by large margins.
• That 200 tokens per second is what you will see. That figure comes from a simulated replay in a blog post.
• That the benchmark numbers for external models (GPT-6 Luna, Claude Haiku 5.5, the open-weight rows) are accurate. Cognition ran these itself and did not publish harness details in the post.
• That "fully American" means anything for capability. It is a procurement and policy statement.

Frequently asked questions

What is SWE-2 Flash?

A coding model from Cognition, released November 4, 2026, with 276B total and 12B active parameters, built on Thinking Machines' Inkling-Small base. It is positioned for routine coding agent work at one-fifth the price and 3x the speed of Cognition's flagship SWE-2, per the vendor.

How do I access SWE-2 Flash?

Through Devin Desktop and the Devin CLI, per Cognition's announcement. The Devin Desktop page offers a macOS download and describes managing agents locally and in the cloud. There is no public API, open weights, or other surface mentioned in the announcement.

Is SWE-2 Flash better than SWE-2?

On Cognition's own table, it depends on the eval. It edges SWE-2 on FrontierCode 2.0 (24.0 vs 23.8) and nearly ties it on SWE-Bench Pro v2 Hard (89.9 vs 90.6), but loses clearly on DeepSWE 1.1 (60.1 vs 73.0) and Terminal-Bench 4.0 (20.0 vs 27.3). The value argument is cost and speed on routine tasks, not across-the-board superiority.

What is mid-training and why does it matter here?

Mid-training is an additional training stage between the base model and reinforcement learning. Cognition says this is the first time it added one, and credits it with moving Terminal-Bench 4.0 from 0.9 to 20.0 and FrontierCode 2.0 from 3.2 to 24.0 over the base model. The data mix was chosen with token efficiency plots measuring tokens needed to hit a target validation loss.

Are the SWE-2 Flash weights open?

No. Cognition says it is building a program to share weights securely with trusted industry partners and research labs, citing dual-use cyber risk, with details promised in the coming weeks. Until then, the model runs only inside Devin.

Should I trust Cognition's benchmark table?

Treat it as vendor-reported. Cognition ran the evals itself, the external model numbers are not independently verified in the post, and the 200 vs 60 tokens per second comparison is a simulated replay. The table is unusually informative for a launch post because it shows losses as well as wins, but it is still the seller's scoreboard.

What is SWE-2 Flash good for, practically?

Per Cognition, the routine workload that dominates coding agent volume: code review, workflow automation, feature development, and documentation updates. The vendor's own numbers suggest keeping a larger model for hard terminal-heavy tasks.

Sources

• Introducing SWE-2 Flash | Cognition - read November 5, 2026. All benchmark numbers, training details, pricing and speed claims, and the weights-sharing plan come from this post.
• Devin Desktop - read November 5, 2026. Confirms the Devin Desktop product surface.
• The Code newsletter, November 5, 2026 issue - where I first saw the release. Its summary matches the Cognition post.

Not read: Devin pricing pages, the Devin CLI documentation, and any independent benchmark replication. Claims above about price ratios, speed, and eval scores are Cognition's, not mine.