# DeepSeek releases V4.1-Flash with 1M context and MIT license

Digest AI · Generative AI & Models · published 2026-09-10T12:40:51Z · updated 2026-09-14T12:00:09Z

Canonical: https://digestai.news/story/deepseek-v4-1-flash-reduces-ai-agent-memory-usage

## Summary

DeepSeek AI has released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model designed to address the memory bottlenecks of long-horizon agent workloads. The model features a 552B parameter backbone with 196B additional Engram parameters and supports a 1M-token context window. By activating only 8B parameters during prefill and 16B during decode, it significantly reduces computational overhead compared to previous iterations.

The release introduces several architectural innovations, including a Causal Encoder-Decoder structure that halves prefill compute and Compressed Sparse Attention 2 (CSA2), which reuses KV cache data across layers. These optimizations reduce the global KV cache footprint to 890 bytes per token, roughly 437 times smaller than DeepSeek-V1. The model also utilizes FP4 quantization for the main KV cache, further cutting storage requirements.

DeepSeek-V4.1-Flash is available under an MIT license with open weights on Hugging Face, supporting deployment via vLLM, SGLang, and Transformers. The team reports that the model outperforms competitors like Opus-5 and GPT-5.6 Sol on specific agentic benchmarks, such as Terminal-Bench 2.1 and DeepSWE v1.1, while maintaining efficient resource usage for enterprise and research applications.

## Key points

- DeepSeek-V4.1-Flash features 552B parameters and a 1M-token context window with MIT-licensed open weights.
- New CSA2 architecture and FP4 quantization reduce KV cache footprint to 890 bytes per token.
- Model outperforms Opus-5 and GPT-5.6 Sol on Terminal-Bench 2.1 and DeepSWE v1.1 agentic tasks.

## Why it matters

This release lowers the hardware barrier for deploying million-token context models, enabling cost-effective long-horizon agents. The MIT license and open weights accelerate adoption in enterprise and research settings, challenging proprietary closed-source alternatives.

## Sources

1. [DeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse](https://marktechpost.com/2026/09/10/deepseek-ai-released-deepseek-v4-1-flash-with-1m-context-fp4-kv-cache-and-cross-layer-attention-reuse) (MarkTechPost, 2026-09-10)
2. [Why DeepSeek-V4.1-Flash Is Such an Exciting Open Model Release](https://kdnuggets.com/why-deepseek-v4-1-flash-is-such-an-exciting-open-model-release) (KDnuggets, 2026-09-14)
3. [DeepSeek V4.1 Flash at 41 tok/s on 8× A40 on TensorSharp](https://github.com/zhongkaifu/TensorSharp) (github.com, 2026-09-12, primary source)
4. [Baseten Adds DeepSeek-V4.1-Flash to Model APIs With 1M-Token Context](https://unite.ai/baseten-adds-deepseek-v4-1-flash-to-model-apis-with-1m-token-context) (Unite.AI, 2026-09-11)
5. [SK Hynix, Samsung Shares Dip In Korea After DeepSeek Debuts AI Tech That Uses Less Memory – Micron, SanDisk Hold Up](https://finance.yahoo.com/technology/ai/articles/sk-hynix-samsung-shares-dip-071811226.html) (finance.yahoo.com, 2026-09-11)
6. [New Deepseek model V4.1-Flash cuts memory needs for AI agents](https://the-decoder.com/new-deepseek-model-v4-1-flash-cuts-memory-needs-for-ai-agents) (The Decoder, 2026-09-10)

Part of the developing story: [DeepSeek Launches Open Long Context Agent Model](https://digestai.news/thread/deepseek-v4-1-flash-reduces-ai-agent-memory-usage) (3 stories)

## Cite

Digest AI, "DeepSeek releases V4.1-Flash with 1M context and MIT license", 10 September 2026, https://digestai.news/story/deepseek-v4-1-flash-reduces-ai-agent-memory-usage

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/deepseek-v4-1-flash-reduces-ai-agent-memory-usage.json