DeepSeek V4.1 Flash, Devin's SWE-2 and a petabyte of DNA predictions

A week where open models and a coding product caught up to the flagships on cost, and where biology got a big data release.

DeepSeek V4.1 Flash

DeepSeek released V4.1 Flash on 10 September under the MIT license. It is a 552B-parameter mixture-of-experts model and the smallest member of a new architecture family, with native image understanding and a 1M-token context. The unusual part is the design: a Causal Encoder-Decoder that activates only 8B parameters when reading input and 16B when writing output. DeepSeek says a smaller KV cache is the main saving, and claims results ahead of its earlier flagship V4-Pro on its own benchmarks. API output can run to 384K tokens.

SWE-2 from Cognition

The same day Cognition introduced SWE-2, the coding model inside Devin. By the company's account it reaches 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1, while costing 64% less. It was post-trained from the Kimi K3 2.8T base, scaled reinforcement learning into the multi-trillion-parameter range, and uses an algorithm that trains every reasoning-effort level in one run. Those are Cognition's numbers on a benchmark Cognition publishes, so keep that in mind.

A genome atlas

DeepMind released the AlphaGenome Atlas on 8 September. It holds precomputed predictions for about 9 billion possible single-letter variants of the human genome, around 1 petabyte, more than 30 times the size of the AlphaFold Database, and free for academic use. For each variant it gives predicted effects across hundreds of human and mouse cell types, plus a single impact score. DeepMind describes using it to spot a rare-disease variant in the DNM1 gene.

Images and graphics

OpenAI released GPT Image 2.5 Flare and Sunburst on 8 September, with Flare as the fast tier. QuiverAI released Arrow 2 on 7 September, a vector model that outputs SVG. Nex AGI opened Nex-N2.5-Max, a 1.6T agentic model under Apache 2.0.

One more paper: NeoHorse-1 describes a route to recursive self-improvement through agentic post-training.

What I'd watch next

Whether DeepSeek's encoder-decoder trick shows up in the next large release, and what else gets post-trained from open bases like Kimi K3.