Last updated: November 5, 2026
SWE-2 Flash, announced by Cognition on November 4, 2026, is a 276B-A12B coding model available in Devin Desktop and CLI at one-fifth the price and about 3x the speed of SWE-2. Cognition says it is free through November 30 for Pro, Max and Teams, and plans controlled weight sharing later.
Contents
- What’s new at a glance
- Where to run it and promotional pricing
- Mid-training, RL and reported benchmarks
- SWE-2 Flash compared with SWE-2 and peers
- How to select SWE-2 Flash in Desktop and CLI
- Devin plans that matter for model access
- Limits, weights and what to verify
- Frequently asked questions
SWE-2 Flash is Cognition’s efficiency coding model for agent work such as automations, code review, feature development and documentation updates. The official Cognition announcement was published on November 4, 2026. It sits beside the larger SWE-2 model inside Devin; see our Devin Desktop 3.11 guide.
What’s new at a glance
| Item | Detail | Where it shows up |
|---|---|---|
| Availability | Devin Desktop and Devin CLI on day one | Select it in the model picker |
| Size | 276B total / 12B active (276B-A12B) | Based on Inkling-Small |
| Speed claim | About 3x SWE-2 generation speed | ~200 tok/s vs ~60 tok/s in Cognition’s demo |
| Price claim | About one-fifth the price of SWE-2 | Exact rates shown in the model selector |
| Promo | Free through November 30, 2026 | Pro, Max and Teams; Desktop and CLI |
| Training change | Mid-training stage before RL | First time in Cognition’s SWE recipe |
| Weights | Not public yet | Secure partner and research program planned |
Where to run it and promotional pricing
According to the SWE-2 Flash launch post, the model is available today in Devin Desktop and Devin CLI. Cognition’s accompanying X thread states that SWE-2 Flash is free until the end of November for all Pro, Max and Teams users on those two surfaces. The blog CTA repeats “Free through 11/30.”
Treat the promo as surface-specific. Devin’s Desktop models docs note that the SWE-2 free window for Desktop and CLI does not apply to cloud sessions. Confirm Flash eligibility in the model selector before a long job.
Cognition does not publish a separate public per-token card for Flash on the launch page. It states a relative price of about one-fifth of SWE-2. For context, the same models documentation lists SWE-2 at $3.00 per 1M input tokens, $15.00 per 1M output tokens and $0.30 per 1M cache input tokens after its own promotional window. Use the live selector for Flash’s current line item.
Mid-training, RL and reported benchmarks
Cognition says this is the first SWE release that adds a large mid-training stage before reinforcement learning. The team started from Inkling-Small (276B total, 12B active), curated held-out coding validation data, compared data mixes with token-efficiency plots, then trained on a few trillion tokens of scientific articles, blogs, code and other sources in three stages that reserved the scarcest high-quality data for last.
Reinforcement learning followed, with audited instructions, environments and verifiers plus new task types aimed at the skills Devin needs. On Cognition’s table, mid-training plus RL lifted Terminal-Bench 4.0 from 0.9% on Inkling-Small to 20.0% on SWE-2 Flash, and FrontierCode 2.0 from 3.2% to 24.0%. Other reported Flash scores include DeepSWE 1.1 at 60.1, Terminal-Bench 2.1 at 89.2, SWE-Atlas QA at 50.0, SWE-Bench Pro v2 Hard at 89.9 and ExploitBench at 34.8 (up from 11.3 on the base).
These figures are Cognition-reported. We did not read independent reproductions for this guide.
SWE-2 Flash compared with SWE-2 and peers
| Model | Size (vendor) | Speed note | Cost note |
|---|---|---|---|
| SWE-2 Flash | 276B-A12B | ~200 tok/s (vendor demo) | ~1/5 of SWE-2 (vendor claim); promo free to Nov 30 |
| SWE-2 | 2.8T-A104B | ~60 tok/s (same demo) | List $3 / $15 / $0.30 per 1M after promo |
| Inkling-Small | 276B-A12B | — | Base model before Cognition mid-training and RL |
| Muse Spark Open | 816B-A42B | — | American open-weight peer on Cognition’s table |
| DeepSeek V4.1 Flash 1017 | 763B-A16B | — | Stronger on several vendor rows; different stack |
On Cognition’s FrontierCode 2.0 row, Flash (24.0) edges SWE-2 (23.8) while trailing Claude Haiku 5.5 (27.7). On Terminal-Bench 4.0, Flash (20.0) trails SWE-2 (27.3) and several larger Flash-class models. The pitch is cost and latency for routine agent turns, not winning every benchmark. For related open-weight and Flash-class coverage, see our Muse Spark Open guide and DeepSeek V4.1 Flash guide.
How to select SWE-2 Flash in Desktop and CLI
- Update or install Devin Desktop, or install Devin CLI from the Devin CLI page (curl, Homebrew or PowerShell).
- Sign in on a Pro, Max or Teams plan if you want the November free window described by Cognition.
- In Desktop, open the model dropdown under the chat or agent input and choose SWE-2 Flash. Devin’s AI Models docs say the list depends on your plan and team allowlist.
- In CLI, run
/modelwith no argument to open the selector, or start withdevin --modelplus the Flash identifier shown in your picker. The CLI models docs also cover pinned models and config defaults in~/.config/devin/config.json. - Run a short, reversible task first and check the usage meter so you know whether the promo or list pricing is applying.
- For cheaper frontier pairing on other work, try Fusion, which Devin recommends as a lead-plus-sidekick setup.
Other terminal agents with recent guides include Codex CLI 0.166 and Mistral Vibe 3.1.
Devin plans that matter for model access
The Devin pricing page currently lists Free at $0 with light agent quota and limited model availability, Pro at $20 per month with full model availability, free SWE-2 use, Devin Cloud access and extra usage at API pricing, Max at $200 per month with higher quotas, and Teams at $80 per month plus $40 per full seat. Free-tier Desktop users are limited to SWE-1.6 in the models docs, so Flash access should be verified on a paid plan during the promo.
Paid plans include a usage allowance that refreshes daily and weekly. Cost per message still depends on the model, task size and reasoning effort after promotions end.
Limits, weights and what to verify
- Weights are not downloadable today; Cognition is designing a controlled sharing program rather than a public dump.
- Benchmark numbers above are Cognition-reported and may use different harness settings than peer blogs.
- Exact Flash token prices after November 30 are not fixed on the launch page; read them from the selector.
- Team admins can allowlist models, so Flash may be missing until an admin enables it.
- Do not assume Devin Cloud sessions inherit the Desktop/CLI Flash promo without checking billing for that surface.
Frequently asked questions
What is SWE-2 Flash?
SWE-2 Flash is Cognition’s fast coding model, announced November 4, 2026. It is a 276B-parameter Mixture-of-Experts model with about 12B active parameters, based on Thinking Machines’ Inkling-Small, and it runs in Devin Desktop and Devin CLI.
How do I use SWE-2 Flash in Devin?
In Devin Desktop, open the model selector under the chat or agent input and choose SWE-2 Flash. In Devin CLI, run /model with no argument to open the picker, or pass –model when you start a session. Confirm the name in the live selector, because model catalogs change often.
Is SWE-2 Flash free?
Cognition’s announcement page and X thread say SWE-2 Flash is free through November 30, 2026 for Pro, Max and Teams users in Devin Desktop and Devin CLI. After the promo, check the in-app model selector for current rates. Devin Cloud sessions are a separate product surface.
How does SWE-2 Flash compare with SWE-2?
Cognition says SWE-2 Flash costs about one-fifth as much and runs about three times as fast as SWE-2, citing roughly 200 tokens per second versus 60 on a Terminal-Bench replay. SWE-2 is the larger 2.8T-A104B sibling; Flash is the 276B-A12B efficiency model for everyday agent work.
Can I download the SWE-2 Flash weights?
Not publicly. Cognition says it is building a program to share weights securely with trusted industry partners and research labs because of cyber-risk concerns, and that more details will follow.
What Devin plans can access frontier models?
Devin’s pricing page lists Free at $0 with limited model availability, Pro at $20 per month with full model availability and free SWE-2 use, Max at $200 per month with higher quotas, and Teams at $80 per month plus $40 per full seat.
Sources
- Introducing SWE-2 Flash (Cognition)
- AI Models in Devin Desktop (Devin Docs)
- Models in Devin CLI (Devin Docs)
- Devin plans and pricing
- Devin CLI
Image: Official SWE-2 Flash announcement cover art from Cognition at cognition.com/images/swe-2-flash/cover.jpg, used for editorial coverage of the November 4, 2026 release. Cognition and Devin are trademarks of Cognition; FreeAnswers is not affiliated with Cognition.