DeepSeek V4.1 Flash Releases with CSA2 KV Sharing and 40 Intelligence Index Score

DeepSeek V4.1 Flash is a 552B-parameter multimodal Mixture-of-Experts model with a Causal Encoder-Decoder architecture that activates 8B parameters per token on input and 16B on output.

The model launched on September 10, 2026, with open weights released under an MIT license on Hugging Face and immediate availability through the DeepSeek API under the identifier deepseek-flash. Native support for image and text inputs distinguishes the release from prior text-only offerings in the series.

Background on DeepSeek Model Evolution

DeepSeek previously positioned the V4 Pro as its flagship offering for high-performance inference tasks. Competitive pressure from efficiency-focused alternatives prompted the development of a lighter yet more capable successor that reduces memory and compute demands without sacrificing benchmark results.

Industry observers had noted rising inference costs associated with large KV caches in earlier generations. The new design directly targets those bottlenecks through shared attention mechanisms and conditional memory structures that reuse computations across layers and tokens.

Key Architectural Innovations

The Causal Encoder-Decoder structure separates prefill and decode phases, limiting active parameters to 8 billion during input processing and 16 billion during generation. This split reduces overall resource consumption while preserving output quality on complex multimodal queries.

Compressed Sparse Attention 2 implements cross-layer KV sharing and index compression. The approach stores only essential key-value pairs and reuses them, yielding the reported reduction in cache size. Engram conditional memory adds a 196 billion parameter store that activates selectively based on context relevance.

Single-Pass mHC and DSpark speculative decoding further accelerate throughput. DSpark generates candidate tokens in parallel and verifies them in a single forward pass, cutting latency on long sequences.

Technical Specifications Overview

Total backbone parameters reach 552 billion under a Mixture-of-Experts routing scheme. Context length extends to one million tokens, enabling processing of lengthy documents or extended conversation histories without truncation.

Comparison of key specifications between DeepSeek V4.1 Flash and its predecessor.

Aspect	DeepSeek V4.1 Flash	DeepSeek V4 Pro
Total Parameters	552B MoE	Higher active count
Active Parameters	8B input / 16B output	Full activation
KV Cache Size	890 bytes per token	4x larger footprint
Intelligence Index	40	Lower reported score
Context Length	1M tokens	Shorter limit
License	MIT open weights	Proprietary elements

Performance Metrics and Efficiency Gains

Global KV cache footprint stands at 890 bytes per token according to the model card. This figure represents one quarter of prior HBM requirements and one eighth of SSD usage, delivering the fourfold cost reduction cited in release materials.

We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens.DeepSeek-AI

Market and Stakeholder Implications

The combination of lower per-token costs and higher benchmark scores positions the model as a direct replacement for enterprise workloads previously served by V4 Pro. Developers gain access to multimodal capabilities without proportional increases in infrastructure spend.

SGLang integration on day zero includes kernel optimizations that reach 873 tokens per second on compatible hardware. This throughput supports real-time agentic applications and high-volume batch processing.

1. Reduced KV cache footprint to 890 bytes per token lowers operational expenses.
2. Multimodal native support expands use cases to image understanding tasks.
3. MIT license enables unrestricted fine-tuning and deployment by third parties.
4. Phased V4 Pro retirement from September 14, 2026, forces migration planning for existing users.

Expert Reactions and Benchmark Context

Independent evaluations from Artificial Analysis confirm the 40-point Intelligence Index result under maximum-effort reasoning settings. The score reflects strong performance across reasoning, knowledge and agentic task categories.

The announcement also states that multi-party testing showed superior total runtime and cost metrics compared with the outgoing V4 Pro, prompting the retirement timeline.

Future Roadmap and Next Steps

Continued scaling of the new Causal Encoder-Decoder series is expected, with additional variants potentially expanding parameter counts while retaining efficiency gains. Integration with additional inference frameworks beyond SGLang will likely follow.

The shift away from V4 Pro signals a strategic emphasis on memory-efficient designs that maintain frontier-level intelligence scores at reduced price points.