DeepSeek Ships V4.1-Flash With Novel Encoder-Decoder Architecture, Cuts KV Cache to 1/8 of Predecessor
DeepSeek released V4.1-Flash, a 763B-parameter model built on a new causal encoder-decoder architecture that splits 8B active parameters for prefill and 16B for decode. The model adds native vision support, a 1M-token context window, and shrinks KV cache footprint to roughly 1/8 of DeepSeek V4 Flash, while retiring V4 Pro.
DeepSeek released V4.1-Flash on September 10, 2026, introducing a new causal encoder-decoder architecture that the company says replaces its previous V4 Pro line entirely. Despite the minor-sounding version number, researchers including Sebastian Raschka argue the architectural shift is substantial enough that it "should have been called DeepSeek V5."
What changed
V4.1-Flash has 763B total parameters but uses a split active-parameter design: 8B active parameters for prefill (input processing) and 16B active parameters for decode (output generation), according to Artificial Analysis. That notation—763B-P8B-D16B—reflects roughly 1-2% sparsity, extending the mixture-of-experts convention to describe the input/output split explicitly.
The model pairs this with a technique DeepSeek calls Sliding-Window Attention Bounded Replay, which according to the company's technical report reduces KV cache footprint to as little as 1/8 that of DeepSeek V4 Flash. The result, according to DeepSeek, is faster inference and higher throughput for long-running agent workloads that depend on extended context retention.
V4.1-Flash also adds native vision support, making it multimodal without a separate model release. Analyst Stochastic Chasm noted the multimodal implementation is not architecturally novel—DeepSeek uses 3x3 pixel unshuffle rather than the more common 2x2 approach and largely lets the existing backbone process visual tokens—but flagged it as a meaningful divergence from vision encoders used by peers like Kimi K3.
Benchmarks and pricing
Artificial Analysis reported V4.1-Flash scores 40 on its Intelligence Index, placing it just below GLM-5.3-Flash but above DeepSeek's own V4 Pro 0813—despite V4.1-Flash costing substantially less to run. Vals AI named it the new #1 open-weight model on the Vals Index, ahead of Kimi K3, running its evaluation suite at just $0.30 per test using 1M context, 384 max output tokens, temperature 1, and high reasoning effort.
Pricing is $0.30 per 1M input tokens and $1.20 per 1M output tokens, with cached input priced at $0.006 per 1M tokens and an additional 50% discount during off-peak hours, according to Artificial Analysis. The model ships under an MIT license with a 1M-token context window and text-plus-image input, available first through DeepSeek's own API with US availability.
Baseten confirmed day-0 support, describing the model as smarter, faster, and more efficient than V4 Pro 0813, with zero-data-retention (ZDR) guarantees for US customers. Ollama began rolling out access to Max and Team accounts, later expanding to Pro plan subscribers.
Post-training philosophy
DeepSeek's commentary accompanying the release suggests a shift in research priorities: the company reportedly stated that, at this stage, improving data quality delivers greater returns than developing new post-training algorithms—a notable admission from a lab that built its reputation on algorithmic innovations like GRPO and R1-style reinforcement learning.
What this means
DeepSeek's decision to bury a genuine architectural departure under a "4.1" label is unusual for a lab that has otherwise used version numbers to signal major jumps. The practical story here is efficiency: an 8B/16B active-parameter split combined with aggressive KV cache compression targets the exact bottleneck that makes long-context agentic workloads expensive—cache memory and prefill latency—rather than chasing raw benchmark leaderboard positions. A 40 on the Artificial Analysis Intelligence Index is not class-leading, but the price-to-capability ratio, MIT licensing, and 1M context window make this a serious option for teams building cost-sensitive, long-running agent infrastructure rather than one-shot chat applications. Whether the encoder-decoder approach generalizes to larger DeepSeek models, as the company suggests it plans to scale, will be the real test of whether this becomes an industry-wide pattern or stays a DeepSeek-specific efficiency play.
Related Articles
InclusionAI Releases Ling 3.0 Flash VL, Adding Vision to Its 124B MoE Model
InclusionAI has released Ling 3.0 Flash VL, a vision-language extension of its 124B total-parameter, 5.5B active Mixture-of-Experts model. The model adds native image and video understanding, supports a 131K token context window, and is priced at $0.06 per 1M input tokens and $0.18 per 1M output tokens via OpenRouter.
DeepSeek V4.1-Flash Cuts KV Cache Memory by Up to 8x, Matches Opus 5 on Coding Benchmark
DeepSeek released V4.1-Flash, a 552-billion-parameter model built to slash the memory overhead of long-context AI agents. The model cuts GPU cache needs to roughly a quarter of its predecessor's and matches closed models from OpenAI and Anthropic on select coding benchmarks.
DeepSeek Launches V4.1 Flash: Low-Cost MoE Model Claims to Beat V4 Pro
DeepSeek has released V4.1 Flash, a sparse mixture-of-experts model priced at $0.30 per 1M input tokens and $1.20 per 1M output tokens with a 1 million token context window. DeepSeek claims the model exceeds the larger V4 Pro on performance, speed, and task completion time.
DeepSeek Releases V4.1-Flash: 552B MoE Model Cuts KV Cache to 890 Bytes Per Token
DeepSeek has released V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts model supporting 1M-token context and activating only 8B parameters during prefill. The model uses a new Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 to cut global KV cache to 890 bytes per token, roughly a quarter of its predecessor.
Comments
Loading...