Chinese artificial intelligence startup DeepSeek on Thursday launched DeepSeek-V4.1-Flash, positioning the system as the smallest model within its newly established architecture family. According to a formal company statement, the release is designed to deliver enhanced capability, faster inference speeds, higher throughput, and scalable foundations for larger future models. The rollout coincides with preparations by the Beijing-based startup for an initial public offering on Shanghai's technology-focused STAR Market, as reported by Reuters.
The multimodal Mixture-of-Experts model incorporates 552 billion backbone parameters alongside 196 billion additional Engram parameters, supporting a comprehensive 1-million-token context window. During operation, the architecture activates 8 billion parameters per token during the prefill phase and scales to 16 billion parameters during the decoding stage. Marktechpost reporting notes that the system achieves a global KV cache footprint of 890 bytes per token, which represents approximately one-quarter of the cache footprint required by the preceding V4-Flash model and roughly 437 times smaller than DeepSeek-V1.
Causal Encoder-Decoder Architecture Halves Prefill Computation
The underlying 40-layer backbone is structurally divided into a 20-layer causal encoder and a 20-layer decoder. Inspired by the YOCO design framework, the decoder avoids computing its own global key-value cache directly. Instead, per-layer projection weights derive the necessary values from the final hidden state of the encoder. Consequently, prompt tokens terminate their computation at the encoder stage, which nearly halves the required prefill compute resources. Sliding-window attention featuring a 128-token window continues to operate across every layer, enabling decoder sliding-window attention states to be rebuilt by replaying only the terminal 128 prompt tokens through a mechanism termed Decoder SWA Bounded Replay.
To manage cache size effectively along the layer axis, DeepSeek-V4.1-Flash implements pure Compressed Sparse Attention 2, replacing the mixed approach utilized in earlier versions. Each compressed sparse attention layer is statically assigned to one of three distinct operational modes known as Full, Reindex, and Reuse. The 18 encoder layers deploy a compression ratio of two across three groups of six layers, while the 20 decoder layers utilize a ratio of one across five groups of four layers. A hierarchical sparse indexer integrated within the decoder enables the Full layer to construct a candidate pool comprising up to 16,384 positions, allowing subsequent reindex layers to evaluate a bounded set rather than scanning the entire context window.
Quantization and Memory Optimizations Drive Deployment Efficiency
The main key-value cache undergoes quantization to E2M1 format utilizing a single E4M3 scale per 16 channels, adhering to NVFP4 specifications without a global scale. Introduced via quantization-aware training during post-training phases, this mechanism nearly halves storage requirements compared to the FP8 cache used in the V4 iteration. At the deployment level, sliding-window attention key-value caches are no longer persisted to solid-state drives, residing instead within a distributed memory pool carved from 10 percent of host dynamic random-access memory with a time-to-live duration of several minutes, whereas the global key-value cache maintains a guaranteed 72-hour lifetime.
Additional structural enhancements include single-pass multi-input hidden coefficient shifting to fuse mega kernel operations and halve activation memory traffic, conditional Engram memory modules positioned at layers 1 and 14, DSpark speculative decoding trained following pre-training with a frozen backbone, and head-wise Muon optimization. Single-token decoding floating-point operations increase by only one-quarter when context lengths expand from 4,000 tokens to 1 million tokens. Open-source weights ship under an MIT license across vLLM, SGLang, and Transformers deployment paths on Hugging Face, alongside a public reasoning API featuring multiple tiers.
Benchmark Performance and Open Weights Availability
Pre-training regimens encompassed 45 trillion multimodal tokens maintained at a 7-to-1 text-to-multimodal ratio. Sparse attention mechanisms were trained from scratch at a 64,000 sequence length without employing a dense warmup phase, while context expansion to 1 million tokens occurred at 34 trillion tokens. The foundational model matches DeepSeek-V4-Pro-Base capabilities across world knowledge and coding benchmarks while consuming one-third of total parameters and one-quarter of activated parameters.
Post-training relied on large-scale synthesis of verifiable agent tasks, reinforcement learning across heterogeneous execution scaffolds including Claude Code, Codex, OpenCode, Pi, mini-SWE, and DeepSeek Harness, alongside on-policy distillation from more than 40 teacher models. Max-effort evaluation results place the model competitively against proprietary systems, recording 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, 31.2 on Terminal-Bench 4.0, 54.8 on Automation-Bench, 90.9 on GPQA Diamond, and a Codeforces rating of 3471, according to technical documentation cited by Marktechpost.
Regulatory Scrutiny and Geopolitical Tech Race Context
The release of DeepSeek-V4.1-Flash unfolds amid intensifying global scrutiny regarding artificial intelligence development methodologies and cross-border capability acquisition. The United States cyber defense agency recently issued an advisory accusing prominent Chinese AI laboratories, including DeepSeek and Moonshot AI, of systematically extracting capabilities from American foundational systems through industrial-scale knowledge distillation campaigns. Ministry of Commerce officials in Beijing rejected the allegations as groundless, characterizing the American regulatory stance as an attempt to enforce industrial monopoly under the pretext of restricting distillation practices.
Diplomatic tensions surrounding artificial intelligence are anticipated to feature prominently during upcoming bilateral discussions between Washington and Beijing. Despite geopolitical headwinds, DeepSeek continues to advance its commercial positioning, expanding developer access to its model weights while progressing toward its planned initial public offering on the Shanghai STAR Market. Technical documentation confirms that the startup's latest infrastructure optimizations are engineered specifically to mitigate the memory and bandwidth bottlenecks that typically accompany long-horizon agentic workloads and million-token reasoning tasks.
