DeepSeek V4.1 Flash (luminal)

DeepSeek V4.1 Flash, released September 10, 2026, is the newest Flash model designed for higher capability ceiling, faster inference, and native multimodal visual understanding integrated into a single architecture. The API offers a 1M-token context with up to 384K maximum output, and DeepSeek reports strong benchmark results including 90.9 on GPQA Diamond, 3,471 Codeforces rating, 90.6 on Terminal-Bench 2.1, 74.2% on DeepSWE v1.1, 88.1 on CyberGym, and 65.4 on NL2Repo-Bench. Compared to V4 Flash 0731, coding results improve sharply: Terminal-Bench 2.1 rises from 82.7 to 90.6 and DeepSWE from 54.4 to 74.2. Flash pricing from September 10, 2026 is $0.003 per million cache-hit input tokens, $0.15 per million uncached input, and $0.60 per million output off-peak, with peak rates double. DeepSeek also routes V4 Pro requests to V4.1 Flash at Flash pricing until V4.1 Pro ships.

DeepSeek V4.1 Flash represents a fundamental architectural shift via a Causal Encoder-Decoder design that cuts KV cache requirements approximately 4× relative to V4 Flash and 437× relative to DeepSeek V1. The model uses 552B backbone parameters with only 8B active during prefill and 16B during decode, plus an additional 196B-parameter Engram conditional memory, delivering major cost efficiency for agentic workloads. Key innovations include Compressed Sparse Attention 2 with three modes, FP4 KV cache compression in E2M1 format, and DSpark speculative decoding. Released September 10, 2026 with MIT-licensed weights, V4.1 Flash outperforms V4 Pro (1.6T/49B active) across measured benchmarks despite having one-third the total parameters and one-sixth the active parameters, marking the first time a Flash-tier model entirely replaces a Pro-tier model. Native multimodal vision is integrated via DeepSeek-ViT, and the analysis covers the full pre-training and post-training pipeline verified against official HuggingFace model card and DeepSeek documentation.

DeepSeek-V4.1-Flash ranks 15th on the LLM Stats composite score, earning top-2% standing in Tool Calling (4 of 202) and top-10% in Coding (7 of 275), with average-tier results in Reasoning, Vision, and Math. The model achieves a CodeForces rating of 3471, GPQA Diamond score of 0.91, and Terminal-Bench 2.1 score of 0.91 at maximum reasoning effort. Its blended cost efficiency places it at $0.24 per million tokens with an LLM Stats Score of 51.2. Cost-efficiency positioning shows DeepSeek-V4.1-Flash at $0.24 with a score of 51.2, positioned between GPT OSS 120B at $0.043/28.7 and Claude Opus 5.5 at $4.76/60.3. The model significantly outperforms its predecessor DeepSeek-V4-Flash-0731, which sits at $0.066 with a score of 43.8, demonstrating the architecture's value improvement. Benchmark methodology uses maximum reasoning effort with temperature 1.0 and top-p 0.95.

DeepSeek V4.1 Flash officially launched on September 10, 2026, as the smallest model in DeepSeek's new architecture family, featuring native multimodal visual understanding. It is a 552B-parameter mixture-of-experts model with a Causal-Encoder-Decoder architecture using asymmetric activation of 8B input parameters and 16B output parameters, reducing costs relative to same-size peers. The model ships on the DeepSeek API under the name deepseek-flash with 1M-token context, and DeepSeek has open-sourced weights on Hugging Face alongside a technical report. Performance-wise, V4.1 Flash surpasses DeepSeek V4 Pro and rivals flagship models like GLM 5.3 and Kimi-K3 on benchmarks such as the Agentic Benchmark, while cutting HBM requirements to one-fourth and SSD needs to one-eighth versus V4 Flash. The release also introduced Peak-Valley Pricing effective 12:00 Beijing time on September 10, with off-peak rates at half the peak price. After September 14, 2026, all deepseek-v4-pro requests route to V4.1 Flash at Flash pricing until V4.1 Pro arrives.