--- title: "DeepSeek's V4.1 Flash out-performs its flagship — and rewrites the cache math" date: 2026-10-17 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-deepseek-v41-flash language: en --- # DeepSeek's V4.1 Flash out-performs its flagship — and rewrites the cache math > DeepSeek's V4.1 Flash introduces a 552B-parameter asymmetric architecture with native vision and a sharply smaller cache footprint, and the lab says it will route the older V4 Pro into this cheaper model. A model that costs more to run usually wins the benchmark. DeepSeek just broke that rule. Its smallest member of the V4 family, V4.1 Flash, posts scores above the heavier V4 Pro it is meant to sit beneath. The trick is not more parameters switched on — it is a different shape of model, built to be cheap to serve rather than cheap to brag about. ## What actually shipped DeepSeek released V4.1 Flash on its API on September 10, 2026, under the model name deepseek-flash, with native multimodal support that lets the model take in images as well as text. It is the smallest model in the new V4 architecture family, but DeepSeek positions it as the workhorse: faster, cheaper, and — on its own numbers — smarter than the flagship it complements. The move follows the V4 preview from April and the V4 Pro general-availability release in August, completing a rapid family rollout. The headline spec is a 552B-parameter mixture-of-experts (MoE) design. What matters more is the new "Causal Encoder–Decoder" structure. Instead of running one big network in both directions, the model uses just 8B active parameters to read your input and 16B active parameters to write its output. That asymmetry is the whole point: reading is cheap, writing is where the model spends compute. DeepSeek has kept the broader V4 family under the permissive MIT license, and for V4.1 Flash it said it will work with the open-source community on inference support. The lab framed the release as part of the large model (大模型) race shifting from "who has the biggest model" to "who can serve intelligence cheaply." ## The cache is the real headline For anyone running agents, the cost is rarely the raw compute — it is the memory needed to remember a long conversation. V4.1 Flash cuts that bill hard: DeepSeek says its KV cache needs only one-quarter of the high-bandwidth memory (HBM) and one-eighth of the SSD storage of the previous generation. Why this matters: cache-hit charges account for a large share of agent running costs, because a coding or research agent re-reads the same long context on every step. Compressing the cache directly compresses those charges. DeepSeek also lowered API prices alongside the launch and kept its off-peak rate at 50% of the peak rate, so flexible workloads can schedule themselves into the cheap window. ## The smaller model that beats the bigger one DeepSeek's own benchmark table puts V4.1 Flash ahead of V4 Pro on several axes. Reported figures include GPQA Diamond 90.9, HLE 36.8 (39.1 on the pure-text subset), a Codeforces rating of 3471, Terminal-Bench 2.1 at 90.6, and DeepSWE v1.1 at 74.2. The lab says third parties have placed V4.1 Flash ahead of V4-Pro on performance, cost, speed, and total runtime. Those are vendor-reported numbers, and they describe a model that is deliberately small on active parameters. The claim is not that V4.1 Flash is universally stronger, but that for the agent and vision tasks most developers actually pay for, the cheaper model now does the job. ## The V4 Pro sunset The release also redraws DeepSeek's own lineup. From 04:00 UTC on September 14, 2026, every request sent to deepseek-v4-pro is routed to V4.1 Flash at V4.1-Flash rates, and that routing stays in place until V4.1-Pro launches. In other words, the heavier, more expensive model is being retired into the cheaper one. DeepSeek named WorkBuddy (including CodeBuddy) and OpenCode as official partners that already fully support V4.1 Flash. The message to developers is consistent: move down the stack to the efficient model, and the ecosystem will follow. ## Why an asymmetric decoder matters Most transformers spend roughly equal compute reading and writing. V4.1 Flash splits them. The encoder side, handling input, runs at 8B active parameters; the decoder side, producing output, runs at 16B. For workloads that read a long context and call many tools, the input path dominates token cost, so trimming it pays off more than trimming the output path. This is a structural bet on agent workloads, where a single task can involve thousands of tool calls against a long, persistent context. In that regime, cache size and active-parameter count — not total parameter count — decide the bill. The design signals where DeepSeek thinks the market is going: not bigger models, but models that are cheaper to run at the scale real products demand. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the real shift is that Chinese labs are now competing on serving economics, not just benchmark points. ## Honest limitations - All benchmark scores (GPQA, HLE, Codeforces, Terminal-Bench, DeepSWE) are DeepSeek's own reported figures; we did not re-run them. - The claim that V4.1 Flash "outperforms V4 Pro" is DeepSeek's; independent replication may differ, and the comparison covers specific agent and vision tasks rather than general ability. - Open-weight status: the V4 family is MIT-licensed, and DeepSeek pledged open-source inference support for V4.1 Flash, but we did not confirm downloadable weights on Hugging Face at publication. - The 8B-input / 16B-output active-parameter split and the 1/4-HBM / 1/8-SSD cache figures come from the official release; we did not inspect the model card or reproduce them. ## What readers can do now - If you run agentic or vision workloads on DeepSeek, switch to deepseek-flash and compare cost against the old V4-Flash before the September 14 routing takes effect. - Benchmark V4.1 Flash on your own tasks — vendor scores rarely match your domain, especially on long-context agent traces. - If you self-host, watch Hugging Face for V4.1-Flash weights and confirm the MIT license terms before any production deployment. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-deepseek-v41-flash Free to quote with attribution and a link to the original.