DeepSeek's new V4.1-Flash model cuts memory footprint

DeepSeek just pulled off another efficiency flex that will give infrastructure accountants a headache.

The company released DeepSeek-V4.1-Flash, a new 552B-parameter Mixture-of-Experts model available through both its API and downloadable weights on Hugging Face under an MIT license. Even wilder, DeepSeek is confident enough in the release that it is retiring its former flagship V4-Pro entirely, automatically redirecting those API requests to V4.1-Flash starting September 14.

Asymmetric brains and tiny cache footprints

The real secret sauce here is an unusual architectural shift.

Instead of a conventional decoder-only setup, V4.1-Flash introduces a Causal Encoder-Decoder architecture. It separates the heavy lifting between ingestion and generation, activating only 8B parameters on input and 16B parameters on output.

DeepSeek also took an axe to the memory footprint. Compared to its predecessor, the model requires just one-quarter of the High Bandwidth Memory and one-eighth of the SSD storage for its key-value cache.

For developers building complex agentic workflows, cache-hit fees are often the silent killer on cloud bills. Slashing that memory overhead allows DeepSeek to offer basement-tier pricing: $0.15 per million uncached input tokens and $0.60 per million output tokens outside peak windows, doubling to $0.30 and $1.20 during peak hours.

Punching above its weight in code

When cranked to maximum reasoning effort, DeepSeek claims Flash delivers genuine frontier results in specific domains.

On the DeepSWE v1.1 coding benchmark, Flash posted a score of 74.2, edging out Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. It also notched wins against both on AutomationBench and topped GPT-5.6 Sol on CyberGym.

Yet, the model is far from omnipotent. Flash fell behind both American flagships on Terminal-Bench 4.0. It also scored 90.9 on GPQA Diamond, which actually trails the retired V4-Pro's 92.4. Across the broader Artificial Analysis Intelligence Index, Flash settled at 40, well behind leaders GPT-6 Astra and Claude Fable 5.1, which tied at 53.

Shadowing the launch is an ongoing intellectual property squabble. Anthropic published a report in September alleging DeepSeek bypassed safeguards to harvest Opus reasoning chains, though it stopped short of confirming whether those traces were used to bake V4.1-Flash.

Why it matters

DeepSeek is playing a game of ruthless economics, and we think it is shifting the battlefield in a very smart way.

For builders and technical teams, V4.1-Flash is worth testing immediately as a workhorse model for coding copilots, routine scripting, and high-frequency data pipelines. With MIT-licensed weights, self-hosting is viable for teams with sufficient compute, while API users can slash operational burn by scheduling batch tasks during off-peak windows. The caveat is reliability: if a task demands infallible terminal execution or advanced domain reasoning, the cost of failed runs and retries could erase the initial token savings.

For the wider industry, this release ramps up immense pricing pressure on Western frontier labs. OpenRouter data showed that open-weight Chinese models captured nearly 30% of token volume during peak periods across 2025. By shipping a fast, cheap model that matches top-tier code completion benchmarks, DeepSeek forces rivals to defend why their premium pricing is worth the extra spend when good-enough intelligence is becoming virtually free.