Cognition SWE-2 Flash Runs 3× Faster at One-Fifth the Cost of Its Flagship Coding Model

Cognition released SWE-2 Flash on November 4, adding a faster, dramatically cheaper coding model to its Devin agent platform. Built by post-training Thinking Machines Lab's Inkling-Small — a 276-billion-parameter sparse mixture-of-experts model that keeps only about 12 billion parameters active per token — Flash delivers coding-task throughput of roughly 200 tokens per second, compared with 60 for the SWE-2 flagship, at approximately one-fifth the price. Paid Devin subscribers can use it at no extra charge through November 30 while Cognition evaluates demand.

The release marks more than an efficiency variant. Cognition used SWE-2 Flash as the first production test of large-scale mid-training in its post-training recipe — a training phase inserted between general pretraining and reinforcement learning that the company says meaningfully lifted coding capability before reward shaping began. The same training process raised the model's performance on ExploitBench, a cybersecurity capability benchmark measuring exploit-finding ability, from 11.3% at the base-model stage to 34.8% after Cognition's post-training — a threefold uplift that led the company to restrict public weight access rather than release them openly.

How Sparse Expert Activation Makes Flash Faster and Cheaper Than SWE-2

The speed and cost gap between SWE-2 Flash and its flagship sibling traces directly to their underlying architectures. SWE-2, released September 10, is post-trained from Kimi K3, a model with 2.8 trillion total parameters and roughly 104 billion active per token. SWE-2 Flash, by contrast, is post-trained from Inkling-Small, which uses 256 routed expert subnetworks plus 2 shared experts — but routes only 6 of those 256 through for each token during inference. The result is a model that carries 276 billion parameters total but runs approximately 12 billion per forward pass.

That eightfold-plus difference in active computation — roughly 104B versus 12B parameters activated per token — is the primary mechanism behind Cognition's claimed 3× speed advantage and 1/5 price differential. Inkling-Small was released by Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, under an Apache 2.0 license on July 30. That open licensing allowed Cognition to treat the base model as a foundation for its own post-training work rather than licensing proprietary weights or training from scratch. Cognition confirms the speed differential with a simulation of a Terminal-Bench 4.0 task called gsea-proteomics, in which the per-task wall-clock difference becomes pronounced in long agentic sessions.

Three-Phase Mid-Training Before Reinforcement Learning: A New Step in Cognition's Recipe

Cognition's standard post-training approach relies heavily on reinforcement learning, rewarding models for completing software engineering tasks successfully. For SWE-2 Flash, the company introduced a large-scale mid-training phase before RL for the first time. Mid-training sits between broad pretraining — where a model ingests general text and code at scale — and RL fine-tuning, where task-specific reward signals shape behavior. It allows targeted data exposure at large volume before reward shaping applies.

Cognition staged this mid-training in three phases, drawing from a data mixture that included scientific articles, blogs, and code, totaling a few trillion tokens. Rather than selecting data by downstream task accuracy, Cognition used token efficiency as the selection criterion — measuring how efficiently the model processes training signal — and held out a separate coding dataset to validate data-mixture decisions. The company says task quality and diversity, rather than scale alone, drove the benchmark gains.

Benchmark Scores Require Context: Flash Leads on One of Three, Trails on Two

Cognition reports three benchmark comparisons for SWE-2 Flash against its SWE-2 flagship and the Inkling-Small base model. All results are company-reported; no independent reproduction has been published as of November 5.

On FrontierCode 2.0 — a 132-task benchmark Cognition developed with open-source maintainers, scored on whether model-generated code produces pull requests that merge successfully — Flash scores 24.0% versus SWE-2's 23.8%. The 0.2-point margin is narrow and within what could be benchmark noise. More importantly, both models sit well below the current open leaderboard leaders on FrontierCode 2.0: Claude Opus 5.5 leads at approximately 45.5%, meaning Flash and SWE-2 are both mid-tier rather than frontier performers on this measure.

On Terminal-Bench 4.0, a 66-task benchmark developed by researchers at Stanford and the Laude Institute that tests terminal-based agent performance, Flash scores 20.0% against SWE-2's 27.3%. GPT-6 Astra leads this benchmark. On DeepSWE 1.1, a software engineering repair benchmark, Flash scores 60.1% versus SWE-2's 73.0%. SWE-2 leads by 12.9 points on that test.

The practical picture is a genuine tradeoff: Flash closes most of the gap with its flagship on FrontierCode 2.0 (the narrow-margin category) while running materially faster and cheaper, but it delivers noticeably lower performance on the two benchmarks that measure terminal-agent and repair tasks — precisely the domain where agentic coding work is most demanding.

Cognition states that Flash "outperforms all American open-weight models" on models "three times larger," but that claim's scope is unspecified and cannot be independently verified. It should not be read as a general performance comparison against frontier AI systems.

Why Cognition Is Withholding Public Weights Despite an Open-Weight Base

The sharpest departure in SWE-2 Flash's release structure is what Cognition is not doing: releasing the post-trained weights publicly, even though Inkling-Small's Apache 2.0 license would permit it. The reason is ExploitBench, a benchmark that measures a model's ability to identify and exploit software vulnerabilities.

Inkling-Small, the base model, scores 11.3% on ExploitBench. After Cognition's post-training, SWE-2 Flash scores 34.8% — a threefold capability uplift in exploit-finding that Cognition connects to October 21, 2026 disclosures of PHP and Node.js remote code execution vulnerabilities. The company made the judgment that freely distributable weights at this exploit-capability level create an unacceptable dual-use risk: the same reasoning and code-execution abilities that make Flash effective for autonomous debugging and software repair also make it effective for offensive security research that would be better controlled.

The resolution is a gated partner program: researchers and trusted enterprise partners can apply for access to run Flash on-premise. Devin users access Flash through the existing cloud platform. This model — open base, restricted post-trained derivative, vetted on-prem path — is distinct from both proprietary API-only deployment and fully open weight release. It represents a voluntary capability-triggered governance posture that other AI labs developing high-capability coding systems will likely face in some form as post-training continues to amplify security-relevant abilities.

What This Means for Developers Running Devin at Scale

For organizations already using Devin for high-volume coding tasks, SWE-2 Flash offers a materially different economics calculation. At 1/5 the price with 3× throughput, the model enables either higher task volume at equivalent cost or equivalent volume at substantially lower spend — assuming the benchmark tradeoffs are acceptable for the task profile. FrontierCode 2.0 parity with SWE-2 (the narrow-margin result) suggests Flash handles the class of tasks that benchmark tests reasonably well; the Terminal-Bench and DeepSWE gaps indicate it underperforms on complex, long-horizon terminal and repair work.

Flash is not available as a standalone API — it runs exclusively inside the Devin agent environment, via Devin Desktop or Devin CLI. That deployment constraint matters for teams that want to route inference directly without the Devin abstraction layer. The free-access window for paid subscribers closes November 30; pricing after that date has not been announced.

The Governance Question That Outlasts This Release

The technical story of SWE-2 Flash — a capable, cost-efficient MoE model built through a novel mid-training and RL recipe — will be familiar to practitioners who have watched each generation of sparse-expert coding models lower the cost floor. The more durable development is the gated-weights decision and what it signals.

Cognition is a company with $48 billion in valuation, more than $1 billion in annualized revenue, and a commercial interest in Devin subscriptions. It chose not to release weights that would allow competitors or independent developers to replicate its deployment at zero marginal cost — that decision alone would not be unusual. What is unusual is the specific justification: a measured, benchmark-quantified uplift in exploit-generation capability that the company treated as a threshold requiring weight restriction. The connection to real-world RCE disclosures in October makes the concern concrete rather than theoretical.

As coding AI moves from experiment to critical infrastructure — running in production software pipelines, handling authentication code, reviewing pull requests — the question of when post-training capability makes open weight release inappropriate will only become more contested. SWE-2 Flash is the first well-documented case where a company drew that line at a specific benchmark score and said publicly why.