DeepSeek Ships V4.1-Flash, a 763-Billion-Parameter Model That Cuts KV-Cache Memory to a Quarter of Its Predecessor's

Overview

DeepSeek unveiled an updated version of its cost-optimized Flash model on Thursday, according to The Register. The new DeepSeek-V4.1-Flash grew substantially larger than its predecessor while its developers say new attention and caching techniques cut the memory needed to serve long conversations, as reported by VentureBeat and The Register. At 763 billion parameters, the point release is more than 2.5 times the size of the model it replaces, The Register reported.

What We Know

• Architecture. V4.1-Flash pairs a 552-billion-parameter mixture-of-experts backbone with a separate 196-billion-parameter “Engram” conditional-memory module that is sparsely accessed via token-based lookups rather than loaded for every computation, according to DeepSeek’s own model card on Hugging Face and VentureBeat. The Register describes the same component as 196 billion “N-gram parameters that form what DeepSeek’s developers refer to as a ‘conditional memory module,’” and compares the approach to the Per-Layer Embedding technique Google’s Gemma team originally built to run models on bandwidth-constrained devices like phones, according to The Register.
• Compute activation. The model uses what DeepSeek calls a Causal Encoder-Decoder architecture, splitting its 40 Transformer layers into a 20-layer encoder and a 20-layer decoder. It activates 8 billion parameters per token while reading a prompt (prefill) and 16 billion while generating a response (decode), according to VentureBeat and DeepSeek’s model card.
• Memory footprint. DeepSeek says the new techniques cut the global key-value (KV) cache to 890 bytes per token, about a quarter the size of V4-Flash’s footprint, per VentureBeat. The Register put the same reduction at “between 13 percent and 25 percent of DeepSeek V4 Flash’s requirements,” adding that the change lets V4.1-Flash “support four to eight times as many users in the same KV cache footprint,” according to The Register.
• GPU requirements. Holding the full model’s weights at FP8 would normally require a minimum of 763GB of GPU memory, but because the Engram weights can be offloaded to system RAM instead of GPU memory, DeepSeek can serve the model with around 567GB of GPU memory, The Register reported.
• Pricing. DeepSeek’s own pricing documentation lists off-peak rates of $0.003 per million cache-hit input tokens, $0.15 per million on a cache miss, and $0.60 per million output tokens; peak rates — Monday through Friday, 01:00-04:00 UTC and 06:00-10:00 UTC — double to $0.006, $0.30 and $1.20 respectively, matching figures reported by VentureBeat. VentureBeat notes those off-peak rates undercut OpenAI’s GPT-5.6 Sol ($4 input / $0.40 cache hit / $20 output), Anthropic’s Claude Opus 5 ($5 / $0.50 / $25) and Moonshot AI’s Kimi K3 ($3 / $0.30 / $15).
• Licensing. The open weights are released under the MIT License on Hugging Face for commercial use, according to VentureBeat and DeepSeek’s model card.
• Benchmarks. DeepSeek’s own published results put V4.1-Flash at 74.2 on DeepSWE v1.1, ahead of the 74.0 it reports for Claude Opus 5 and 73.0 for GPT-5.6 Sol, plus 88.1 on CyberGym and 54.8 on AutomationBench. Anthropic’s Opus 5 leads on Terminal-Bench 3.0 (43.3 to 30.0) and Terminal-Bench 4.0 (51.8 to 31.2), and GPT-5.6 Sol leads on GPQA Diamond and SEC-Bench Pro in DeepSeek’s own comparison table, according to VentureBeat.
• Predecessor comparison. DeepSeek’s previous V4-Flash used a 284-billion-parameter backbone with 13 billion active parameters. The new backbone is roughly 94% larger, and while prefill activation fell from 13 billion to 8 billion parameters, generation activation rose from 13 billion to 16 billion, per VentureBeat.
• Migration. DeepSeek has retired V4-Flash and V4-Flash-Vision-Exp, temporarily routing their existing API identifiers to V4.1-Flash. The company says that starting September 14, calls to “deepseek-v4-pro” will also route to V4.1-Flash until a future V4.1-Pro model arrives, according to VentureBeat.
• Industry parallel. The Register reported that Alibaba revealed an experimental 180-billion-parameter model last month, Qwen 3.8-Flash-Next, that pairs a 51-billion-parameter pool of N-gram weights using techniques from the same DeepSeek research, and that Alibaba says this architecture will underpin its next-generation Qwen 4 models.

What We Don’t Know

• DeepSeek’s benchmark figures — including the DeepSWE, CyberGym, AutomationBench and Terminal-Bench scores — are the company’s own reported results rather than independently verified numbers, VentureBeat noted.
• A third-party claim relayed by VentureBeat, that a company called OpenDesign found V4.1-Flash reached “98% of GPT-6 Astra’s quality score at 1.4% of its cost” on its own design-task workload, is, in VentureBeat’s own characterization, “a narrow third-party workload, not a general model evaluation.”
• DeepSeek has acknowledged robustness limits it says it has not fully characterized in the new architecture. VentureBeat reported the company flagged that sparse-attention selection errors and its approximate cache-reconstruction method “could cause capability degradation in untested edge cases,” particularly around long-context retrieval and cache-resumption boundaries, though DeepSeek says it has not observed systematic problems in its own testing so far.
• Some developers reacted skeptically on Hacker News, per VentureBeat. One commenter questioned whether a 552-billion-parameter backbone is “not really flash anymore,” given the added local-deployment hardware demands, while others argued that silently rerouting an existing production identifier like “deepseek-v4-pro” to a different model risks invalidating regression testing for teams built around the old model’s behavior.