DeepSeek V4.1 Flash hits sixth globally days after launch

DeepSeek released V4.1 Flash on September 10, a multimodal Mixture-of-Experts model that carries 552 billion backbone parameters but activates only a small fraction during inference, marking a shift in how large language models balance scale with efficiency. Within three days of launch, the model climbed to sixth place globally in token consumption on the OpenRouter platform, according to data tracker Gelonghui.news.futunn

The model replaces DeepSeek's earlier V4 Flash and V4 Flash Vision offerings, which now route automatically to V4.1 Flash through the API. Starting September 14, all requests to the V4 Pro endpoint are also being temporarily redirected to V4.1 Flash and billed at its lower rate until V4.1 Pro ships.hi-tech+1

Architecture and Efficiency

V4.1 Flash is built on a Causal Encoder-Decoder architecture with 552 billion total parameters, but its Mixture-of-Experts design activates just 8 billion parameters during input processing and 16 billion during output generation. That is a reduction from the 13 billion activated per token in the previous V4 Flash, which had 284 billion total parameters.hackernoon+2

The model supports a context window of up to one million tokens and uses a redesigned KV-cache that stores roughly 890 bytes per token, about one-quarter the memory footprint of V4 Flash. For organizations running large-volume workloads, the reduction in memory requirements may matter more than raw benchmark gains. The model weights are publicly available on Hugging Face under an MIT license.mindstudio+2

Early Adoption and Market Position

Global AI model token consumption reached 127 trillion tokens during the week of September 7–13, up more than 10 percent from the prior week, according to OpenRouter data reported by Gelonghui. Chinese AI models accounted for 61.17 trillion of those tokens, maintaining their lead over U.S. models for the twentieth consecutive week.news.futunn

V4.1 Flash debuted at sixth place with 4.94 trillion tokens consumed in its first three days, while Xiaomi 's MiMo-V2.5 surged 230 percent week over week to claim fifth. Alibaba Cloud ↘ 0.06%, meanwhile, launched its own hosted DeepSeek-V4.1-Flash platform on September 13, offering API access and a Token Plan integrated with tools including Qoder and Codex.gurufocus+1

Benchmark Performance and Limitations

At maximum reasoning effort, V4.1 Flash scored 74.2 on DeepSWE v1.1, 90.6 on Terminal-Bench 2.1, and 88.1 on CyberGym, according to DeepSeek's own evaluations. The model also demonstrated strong multimodal results, including 95.6 on DocVQA and 56.5 on MMMU-Pro.flowtivity+1

However, V4.1 Flash trails DeepSeek's own V4 Pro on several knowledge and reasoning benchmarks, and its long-context LongBench-V2 score of 45.2 falls below the V4 Pro's 51.5. DeepSeek has not published VRAM requirements, inference-speed benchmarks, or batch-size guidance, leaving local deployment planning largely to developers' own testing.hackernoon