Cognition SWE-2 Flash: A 276B-A12B Coding Model Built on Inkling-Small, Free in Devin Through November 30

Cognition has released SWE-2 Flash, a coding model it calls "fully American," built on Thinking Machines' open-weight Inkling-Small. The company announced it on November 4, 2026 in a blog post. It is available now in Devin Desktop and the Devin CLI, free through November 30, at one-fifth the price and three times the speed of SWE-2, Cognition's larger model. The weights are not public. Cognition plans a program to share them with vetted partners and researchers.

The numbers are Cognition's own, and the comparison set is chosen by Cognition. This post covers what was built, what the tables show, what is and is not verified, and who might want to try it.

TL;DR

What Cognition built

The base is Thinking Machines' Inkling-Small, an open-weight mixture-of-experts model with 276B total and 12B active parameters, released in July. Cognition's recipe has two stages. First, a large-scale mid-training stage on a curated mix of a few trillion tokens of scientific articles, blogs, code and other sources. The team tested data mixes by measuring how many tokens each needed to match a baseline's validation loss, and fed the scarcest, highest-quality data late, in three stages. Second, large-scale reinforcement learning. Cognition says most of the gains over SWE-2's RL recipe came from better quality and diversity of RL tasks. It audited instructions, environment setup and verifiers, and added new task types.

Cognition SWE-2 Flash coding model announcement graphic Image: Cognition

The company says this is the first time it added mid-training before RL. Its headline result: Terminal-Bench 4.0 went from 0.9% to 20.0%, and FrontierCode 2.0 from 3.2% to 24.0%, compared with the base model. Cognition describes SWE-2 Flash as better than every American open-weight model on leading evals, including models three times larger.

The benchmark table

Cognition's page includes a table across several models. Values are percentages.

The page also lists Beam 1.0, GLM 5.3 Flash, DeepSeek V4.1 Flash, GPT-6 Luna and Claude Haiku 5.5, and further benchmarks such as Terminal-Bench 2.1, SWE-Bench Pro v2 Hard, SWE-Atlas QA, FrontierCode 1.1 Main and ExploitBench. On FrontierCode 2.0 the page shows 24.0 for SWE-2 Flash against 21.9 for GPT-6 Luna and 27.7 for Claude Haiku 5.5. On Terminal-Bench 4.0 it shows 20.0 against 16.4 and 39.2. Two open models also score higher than SWE-2 Flash on Terminal-Bench 4.0: GLM 5.3 Flash at 41.8 and DeepSeek V4.1 Flash 1017 at 42.3. Both are Chinese models, so they sit outside Cognition's "American open-weight" claim. The pattern is that SWE-2 Flash is competitive with small closed models on some tests and well behind on others. That is a more modest claim than "parity", which is the wording some social posts used.

The page's size row lists SWE-2 Flash and Inkling-Small at 276B-A12B, Muse Spark Open at 816B-A42B, and SWE-2 at 2.8T-A104B. That is how Cognition supports its "even models three times larger" line: its main open comparison, Muse Spark Open, is about three times bigger in total parameters.

What mid-training adds, in plain terms

Most coding models are made in stages. Pretraining teaches general language and code from huge text collections. Post-training then shapes behavior, usually with supervised examples and reinforcement learning, where the model is rewarded for tasks that pass automatic checks. Mid-training sits between the two. It keeps training on the base model's objective, next-token prediction, but on a smaller, higher-quality, more targeted mix. In SWE-2 Flash's case that mix was a few trillion tokens of scientific articles, blogs, code and other sources, chosen with a held-out coding dataset that Cognition built to reflect real software engineering.

The reason this matters is the jump on agent benchmarks. Terminal-Bench 4.0 tests long command-line tasks. A base model that scores 0.9% cannot do them at all, so RL has little to reward. Mid-training gives the model enough skill that RL has something to build on. That is Cognition's own explanation of the gain, and it is plausible, but the page does not separate how much came from mid-training and how much from the improved RL tasks. The company says most RL gains came from task quality and diversity, so the credit is shared.

The validation-loss charts on the page show loss falling at a stable log-linear rate through each of the three stages. That tells you the training was smooth. It does not tell you how the model behaves on your codebase, which only a trial can show.

Why a base model from another lab matters

SWE-2 Flash starts from a model Cognition did not train from scratch. Inkling-Small came from Thinking Machines with open weights, so any team could build on it. Cognition's earlier SWE-2 was built on a different base, and the larger SWE-2 is listed at 2.8T total parameters with 104B active. Using a smaller open base cuts training cost and lets a product company ship a fast model of its own. It also means the quality of the base, and the terms under which it was released, shape what the downstream model can be. That is why the decision not to publish the new weights stands out: the open base made the work possible, but the derived model is closed for now. Readers who follow open-weight policy will want to see whether the partner program stays narrow or widens.

Speed and price

Cognition shows a replay of each model's own run on a Terminal-Bench 4.0 task, at a simulated 200 tokens per second for SWE-2 Flash against 60 for SWE-2. It says SWE-2 Flash costs one-fifth as much. The page does not give a per-token price in the text we could read. For how SWE-2 itself launched and was priced, see Cognition SWE-2 for Devin in September, and for how Devin cut costs on the CLI, see Devin Fusion's 39% cost cut.

The cyber angle

The most interesting paragraph is about origins and cyber. Cognition says enterprises and governments need to host and control deployments on their own infrastructure, and researchers need to study how models behave and can be misused. So it wants to give vetted partners transparent access to weights. It also says it targeted cybersecurity work in training, raising the model's ExploitBench score from 11.3% to 34.8%. Because the same skills could help attackers, "as demonstrated by a recent surge in cyber attacks," it is not releasing weights publicly yet. The weight-sharing program details are promised in a few weeks.

That is a real shift: a company building on an open-weight base chooses not to open the result. It echoes the debate we covered in the American closed AI versus China open weights strategy debate.

Claimed versus verified

Comparisons to Cognition's earlier numbers are also worth a look. In October we examined its FrontierCode 2.0 claim against Opus 5.5, and the same caution applies: the benchmark is one the company helped build.

Should you try it

Because it is free in Devin through November 30, the cost of testing is low. Try it on the work Cognition names: automations, code review, feature development and documentation updates. Compare it against your current model on your own repository, and record cost, speed and the number of fixes you had to make. If you need open weights, wait for the partner program, since the model cannot be downloaded today. If you need on-prem hosting, ask Cognition directly about the program's terms.

What to watch next

• The weight-sharing program details, promised in a few weeks.
• Independent runs of Terminal-Bench 4.0 and FrontierCode 2.0.
• A per-token price once the free period ends.
• Whether Thinking Machines responds, since its open base made this possible. See Inkling's open weights for the original release.

Related reading

• Inkling-Small open weights
• Cognition SWE-2 launch
• Devin Fusion CLI cost cut
• FrontierCode 2.0 claim vs verified
• American closed AI vs China open weights

Figures come from Cognition's November 4, 2026 announcement and may change. Follow @explainx_ai for updates.