Chinese AI lab DeepSeek has released V4.1-Flash, a multimodal open-weight model that pairs a 552-billion-parameter backbone with some of the most aggressive memory compression yet shipped in a frontier-class system. The model went live on Wednesday with an MIT license on Hugging Face and an accompanying technical report, Reuters reported, capping a run of headlines for the Hangzhou-based lab.
The release is more than a routine version bump. DeepSeek has simultaneously retired its flagship V4 Pro tier, and TechNode reported that requests to V4 Pro are now being routed to V4.1-Flash — a striking statement of confidence in a model that carries the "Flash" name yet posts benchmark numbers at or above the model it replaces. For more context on this story, see our ongoing more AI stories.
A Bigger "Flash" With a Smaller Memory Bill
According to the official model card, DeepSeek-V4.1-Flash is a Mixture-of-Experts (MoE) model with 552 billion backbone parameters plus a 196-billion-parameter "Engram" conditional memory component that is sparsely accessed through token-based lookup. Despite the sheer size, the model activates only 8 billion parameters per token during prefill and 16 billion during decode — fewer active parameters than its 284B-parameter predecessor, which ran a constant 13 billion.
The architectural centerpiece is what DeepSeek calls a Causal Encoder-Decoder (CED) design: a 40-layer Transformer split into a 20-layer causal encoder followed by a 20-layer decoder. Because the decoder's global KV cache is projected from the final encoder hidden states rather than accumulated layer by layer, the memory needed to sustain long conversations shrinks dramatically.
The compression numbers are the headline. With FP4 main KV caching in E2M1 format, the global KV cache footprint drops to 890 bytes per token — roughly one quarter of DeepSeek-V4-Flash. A technique called SWA Bounded Replay reconstructs missing attention states by replaying only the most recent window of tokens, cutting the persistent KV cache to about one eighth of the previous Flash model. Per the model card, the reduction is roughly fourfold against V4-Flash and 437-fold against DeepSeek-V1. Additional components include Compressed Sparse Attention 2 (CSA2), which assigns each attention layer one of three static modes, and DSpark speculative decoding for faster generation.
Under the hood, each MoE layer combines one shared expert with 384 routed experts, activating six routed experts per token.
Benchmarks: Ahead of the Retired Flagship
On the base-model evaluations in the technical report, V4.1-Flash edges past the much larger V4 Pro on several key tests: 74.1 on MMLU-Pro versus 73.5, 79.4 on HumanEval versus 76.8, 60.6 on BigCodeBench, and 93.0 on GSM8K. The South China Morning Post reported that DeepSeek says the new model beats Kimi K3 on cyber and coding benchmarks, a notable claim in a Chinese open-model field that also includes Z.ai's GLM-5.3 and Alibaba's Qwen line.
At maximum reasoning effort, the instruct model scores 90.9 on GPQA Diamond — impressive, though still behind the leading frontier systems referenced in DeepSeek's own comparison table, GPT-5.6 Sol at 94.1 and Claude Opus 5.0 at 93.4. Kimi K3 sits at 92.9. The honest read: this is a frontier-adjacent model at open-weight prices, not a clean sweep of the closed labs.
V4.1-Flash is also genuinely multimodal. A DeepSeek-ViT vision encoder, trained from scratch, feeds images through a two-layer projector, and the model was trained on text and images jointly from the start of pre-training. It scores 56.5 on MMMU-Pro and 95.6 on DocVQA.
Built for Agents, Priced for Volume
The design choices make the target audience clear: agentic workloads, where models read enormous contexts and act over long horizons. The model supports context windows of up to one million tokens, was trained on a 45-trillion-token multimodal corpus, and offers a continuously controllable reasoning-effort dial from 1 to 100 that trades inference cost against accuracy. The Decoder's coverage emphasized that the memory savings specifically reduce what it costs to run AI agents — workloads that spend far more time reading input than generating output.
Community reaction has been emphatic. The launch thread on Hacker News drew more than 650 points, and an earlier thread titled "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro" collected nearly 400 more. Commenters who run models locally noted that the Engram parameters can even be offloaded to fast SSDs, opening a path to near-frontier inference on consumer hardware.
Seeking Alpha framed the commercial stakes bluntly, calling it an ultra-low-cost model that presents challenges for US frontier labs — a reminder that the price war in AI inference shows no sign of cooling.
A Deliberate Moment for the Announcement
The timing is telling. A day before the launch, Reuters reported that DeepSeek had tapped CITIC Securities to arrange an IPO on Shanghai's STAR Market, following earlier reporting that the lab's revenue has grown tenfold ahead of a possible listing. Shipping a headline-grabbing open model days later keeps that momentum going.
The release also lands amid heightened scrutiny of Chinese AI firms. Earlier this week, US agencies including the NSA, CISA and FBI named six Chinese AI companies in an advisory about industrial-scale model distillation — a reminder that DeepSeek's technical wins now arrive alongside geopolitical baggage.
None of that dims the engineering story. With V4 Pro retired and its traffic routed to a cheaper, stronger model, DeepSeek has made the clearest possible argument that in 2026 the interesting frontier of AI may not only be who builds the biggest model — but who serves the most capability per gigabyte of memory.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
