DeepSeek V4.1 Flash
A lightweight, fast-inference model in the DeepSeek series belonging to the Flash family. It features lower inference costs and is suited for high-frequency invocation scenarios.
Core Facts
The DeepSeek Flash version is a distilled model optimized for low-latency scenarios, trading some complex reasoning capability for faster response speeds and lower token unit prices.
DeepSeek V4 Flash is an MoE multimodal model with a total parameter count of 284B.
DeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
Timeline (last 90 days)
OpenCode announced that in its $10-per-month Go plan, the model usage value of DeepSeek V4.1 Flash is increased from $60 to $80, with the plan price unchanged.
The separation of the read and write paths is reflected in the computation flow, while the sharing of memory is reflected in state reuse; the two are not contradictory but are the premise that makes this design viable.
Selective reconstruction requires making a judgment about 'which history is relevant to the current query', and this judgment itself incurs additional overhead; whether it is worthwhile depends on whether the computation saved by skipping reconstruction exceeds the judgment cost.
DeepSeek V4.1 Flash lets the read and write paths share the same KV state: the read phase establishes and updates this memory, and the generation phase directly reuses it, rather than starting another cache.
The selective reconstruction mechanism of DeepSeek V4.1 Flash reconstructs only the historical parts that truly affect the current output, leaving the rest unchanged.
The core of DeepSeek V4.1 Flash's architectural adjustments contains three points: a shorter input path, shared attention memory, and a selective reconstruction mechanism.
On Terminal Bench v2.1, Beam's 80.1 trails DeepSeek V4.1 Flash's 90.6; on DeepSWE v1.1, Beam's 44.4 trails GLM-5.3's 61.0 and Kimi K3's 68.0.
When new models released by various labs in September replaced older models in the same product line, GPT-6 Luna fell 56%, GPT-6 Sol fell 50%, DeepSeek V4.1 Flash fell 20%, and Claude Opus 5.5 fell 20%, while the other six remained flat—none increased in price.
DeepSeek V4.1 Flash with the thinking feature disabled did not lose a single scorer response, and the total cost for the entire scan was only $0.14.
V4.1 Flash can deliver a four- to six-fold speed improvement in scenarios such as ultra-long context retrieval, SVG code generation, and large SQL queries.
40 more timeline events
All Facts (20)
The DeepSeek Flash version is a distilled model optimized for low-latency scenarios, trading some complex reasoning capability for faster response speeds and lower token unit prices.
90%VerifiedDeepSeek V4 Flash is an MoE multimodal model with a total parameter count of 284B.
80%VerifiedDeepSeek V4 Flash is a cost-effective inference model released by DeepSeek, optimized for speed and cost.
80%VerifiedDeepSeek V4.1 Flash claims to achieve breakthroughs across three dimensions: capability, speed, and cost.
75%VerifiedV4.1 Flash is natively multimodal in design, capable of directly viewing and understanding image content.
75%VerifiedDeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
75%VerifiedDeepSeek V4.1 Flash scores lower than the previous version on GPQA, HLE, and other benchmarks.
75%VerifiedCompared to the previous-generation V4 Flash, V4.1 Flash reduces the demand for high-bandwidth memory (HBM) to one-quarter and the demand for SSD persistent storage to one-eighth.
75%VerifiedMistral Large 4 scored only 38 points on the Artificial Analysis composite intelligence index, ranking sixth among open-source models after excluding closed-source models.
70%VerifiedV4 Flash's cache size is only 7% of that of V3.2.
70%VerifiedV4.1 Flash adopts a new CED (Causal Encoding) architecture, whose core idea is an asymmetry in computation between input and output.
70%VerifiedV4.1 Flash surpassed Opus 5 on the DeepSWE benchmark by a narrow margin of 0.2 points.
70%VerifiedThe Flash version is positioned as a lighter, faster inference option suited for latency-sensitive or resource-constrained scenarios, while the PRO version targets a higher capability ceiling with greater demands on VRAM and compute
70%VerifiedV4.1 Flash activates 8 billion parameters when processing input and activates 16 billion parameters when generating output.
65%VerifiedCompared to the previous generation Flash, V4.1 Flash's high-bandwidth memory requirements have been reduced to 1/4, and solid-state storage requirements have been reduced to 1/8.
65%VerifiedThe browser operating system interface implemented by DeepSeek V4.1 Flash presents a dynamic desktop with animation effects, features refined popup designs, and Dock bar applications can be invoked normally.
65%VerifiedDeepSeek V4.1 Flash adopts a Causal Encoder-Decoder asymmetric architecture with a total parameter count of 552B.
65%VerifiedCompared to the old Flash model, V4.1 Flash has more comprehensive capabilities and more accurate recognition, but takes longer and costs more per session.
65%VerifiedInterfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.
65%VerifiedV4.1 Flash activates only 8 billion parameters during the input stage and 16 billion parameters during the output stage.
65%