DeepSeek V4.1 Flash

2026-09-10 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V4.1-Flash, a multimodal MoE of 552 billion parameters with a causal encoder-decoder architecture, with 8 billion active parameters on read, 16 on generation, under the MIT license.

### Context V4-Pro had set the reference level of the line. V4.1-Flash is presented as the smallest model of a new architecture family, designed for a higher capability ceiling, faster inference and a move to larger models. It also absorbs V4-Flash and the experimental vision model V4-Flash-Vision-Exp (August 21, 2026). On September 14, 2026, requests to V4-Pro are redirected to V4.1-Flash until the launch of V4.1-Pro. ### What the model brings - A 40-layer causal encoder-decoder architecture: the decoder KV cache is derived from the final state of the encoder, which limits activation to 8 billion parameters on read and 16 on generation. - CSA2 attention with three modes per layer (Full, Reindex, Reuse), which shares sparse attention indices across layers. - A reduction of the KV cache: about 4 times less global cache per token than V4-Flash (model card), i.e. a quarter of the HBM memory and an eighth of the SSD storage (post). - Native vision: a DeepSeek-ViT visual encoder and a projector process images and text from pre-training onward. - A reasoning effort continuously adjustable from 1 to 100. - Auxiliary components: Engram conditional memory (196 billion parameters), DSpark speculative decoding, single-pass mHC. ### Architecture and specifications - Type: multimodal MoE, 552 billion backbone parameters (Hugging Face shows a higher total with the auxiliary modules), 1 shared expert and 384 routed experts per layer, 6 active per token. - Context: 1 M tokens; maximum output of 384K (pricing page). - Input: images and text; output: text. - Pre-training: 45 trillion multimodal tokens, 64K sequences then extension to 1 M. - License: MIT (code and weights), commercial use allowed. - Weights: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash. Technical report published in the repository. - Settings: temperature 1.0, top_p 0.95 or 1.0, output of at least 256K tokens. Self-hosting hardware: not disclosed. ### Cost and access | API `deepseek-flash` (USD per million tokens) | Input, cache hit | Input, cache miss | Output | |---|---|---|---| | Off-peak | 0.003 | 0.15 | 0.6 | | Peak | 0.006 | 0.3 | 1.2 | Effective September 10, 2026 at 04:00 UTC; peak from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays. The identifiers `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are temporarily redirected to this model. ### Benchmarks | Benchmark (max effort) | Opus-5.0 | GPT-5.6 Sol | Kimi K3 | V4-Pro | V4.1-Flash | |---|---|---|---|---|---| | GPQA Diamond | 93.4 | 94.1 | 92.9 | 92.4 | 90.9 | | Humanity's Last Exam | 56.3 | 44.5 | 43.5 | 42.7 | 36.8 | | Terminal-Bench 2.1 | 89.1 | 88.8 | 88.3 | 87.9 | 90.6 | | Terminal-Bench 3.0 | 43.3 | 34.4 | 17.7 | 11.8 | 30.0 | | DeepSWE v1.1 | 74.0 | 73.0 | 67.5 | 62.7 | 74.2 | | CyberGym | not published | 84.5 | 80.0 | 83.3 | 88.1 | | HLE with tools | 63.6 | not published | 59.8 | 60.0 | 63.9 | | AutomationBench | 50.3 | 45.8 | 46.7 | 43.2 | 54.8 | Source: DeepSeek-V4.1-Flash model card on Hugging Face, scores measured and published by the vendor. ### Versus closed models At the time of its release, the closed models compared by the vendor are Opus 5.0 and GPT-5.6 Sol. On agents, V4.1-Flash is ahead of both on Terminal-Bench 2.1 (90.6 versus 89.1 and 88.8), DeepSWE v1.1 (74.2 versus 74.0 and 73.0), CyberGym (88.1 versus 84.5) and AutomationBench (54.8 versus 50.3 and 45.8). It remains behind on GPQA Diamond (90.9 versus 93.4 and 94.1), on Humanity's Last Exam (36.8 versus 56.3 for Opus 5.0) and clearly on Terminal-Bench 3.0 (30.0 versus 43.3). The price gap is not quantified, as the closed models' pricing is not in the source; the vendor lists 0.6 dollars per million output tokens off-peak. The vendor also states that multi-part tests place V4.1-Flash ahead of V4-Pro in performance, cost, speed and total duration. ### Key takeaways V4.1-Flash shows that a smaller, cheaper open model can match or exceed the current closed models on several agent benchmarks. The KV cache reduction directly targets the cost of agents, where the cache dominates the bill.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Multimodal, Agents