--- title: DeepSeek V4.1 Flash released on Hugging Face, replacing... description: DeepSeek V4.1 Flash has launched on Hugging Face as a 552B parameter Mixture-of-Experts model featuring a new Encoder-Decoder architecture, a new pre-training... canonical: https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek V4.1 Flash released on Hugging Face, replacing V4 Pro | daily.dev og:description: DeepSeek V4.1 Flash has launched on Hugging Face as a 552B parameter Mixture-of-Experts model featuring a new Encoder-Decoder architecture, a new pre-training... og:url: https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p og:image: https://api.daily.dev/og/posts/zIKP1p10p.png og:image:alt: DeepSeek V4.1 Flash released on Hugging Face, replacing V4 Pro og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DeepSeek V4.1 Flash released on Hugging Face, replacing V4 Pro **[Collections](https://daily.dev/sources/collections)** · 4 min read · 5 upvotes · 0 comments ## Summary DeepSeek V4.1 Flash has launched on Hugging Face as a 552B parameter Mixture-of-Experts model featuring a new Encoder-Decoder architecture, a new pre-training method, and larger-scale reinforcement learning post-training. It supports native visual understanding and is the smallest model in DeepSeek's new architecture family. DeepSeek says the model outperforms V4 Pro on performance, cost, speed, and compute usage, and as a result V4 Pro is being retired and phased out once V4.1 Flash is fully live. ## Content DeepSeek released V4.1-Flash, a 552B-parameter mixture-of-experts model that only activates 8B parameters during prefill and 16B during decode. Despite the modest-sounding name, several people in the ML community have pointed out this is closer to a major release than a minor update. ## Architecture changes The biggest structural shift is a move to a Causal Encoder-Decoder (CED) design. The decoder's KV cache is built from the encoder's final hidden states rather than from full decoder computation, which cuts memory use significantly. DeepSeek claims V4.1 Flash needs roughly a quarter of the KV cache memory of V4-Flash. A few other pieces worth noting: - **Engram module**: 196B parameters dedicated to n-gram memory lookup rather than computation. About a quarter of the checkpoint is this memory store, which the model queries instead of recomputing. - **Compressed Sparse Attention 2 (CSA2)**: only four layers write compressed KV; the rest of the model shares it. - **FP4 KV cache** with quantization-aware training to reduce storage further. - **Native multimodal support**: vision embeddings from a ViT-based DeepSeek-ViT encoder are passed alongside text tokens. Training used SigLIP loss first, then autoregressive loss for the combined vision+LLM system. - Trained on 45T tokens of multimodal data. The active parameter count (8-16B) makes it unusually cheap to run for its capability level. ## Post-training DeepSeek explicitly said they didn't introduce new post-training algorithms here. Their view is that engineering the data and environment pipeline is currently more valuable than algorithmic novelty. They built a production-grade sandbox platform called DeepSeek Elastic Compute (DSec) to run large-scale agentic training environments, and used the model itself to construct its own training data from internal usage. They also added an explicit effort-level conditioning signal during RL: max (b=100), high (b=75), low (b=50). The final post-training stage used over 40 teacher models. ## Benchmarks Results are mixed depending on what you care about. On coding and cybersecurity benchmarks it matches or beats Claude Opus 5 and GPT-5.6 Sol. On Humanity's Last Exam and ProgramBench it trails. On the Vals Index it's currently the top open-weight model. On the Artificial Analysis Index it scores 40, above V4-Pro but below GLM-5.3-Flash. One practical test: a landing page built with V4.1 Flash cost 2.6 cents versus $1.21 for Claude Fable 5, at comparable quality. Another benchmark showed it reaching 98% of GPT-6 Astra's score at 1.4% of the cost. A retest with thinking enabled (max effort) scored 81.25% on a custom coding benchmark, up from 53.75% with thinking off, and hit roughly 221-235 tokens per second including reasoning tokens. ## Pricing and API routing Prices dropped as much as 32% from August levels. Off-peak API pricing is $0.15 per million input tokens and $0.60 per million output tokens, doubling at peak. Starting September 14 at 04:00 UTC, all V4-Pro traffic routes to V4.1 Flash at flash pricing until V4.1-Pro ships (no date given). Old V4-Flash API names already route to V4.1 Flash. ## Running it locally The model is available on Hugging Face and vLLM supports it from day one on both NVIDIA and AMD GPUs. For local Mac setups, antirez got 15 t/s on a single 128GB M5 Max using SSD streaming in DwarfStar, and 25 t/s with tensor-parallel execution split across two MacBooks via RDMA. It's not really a fit for a single 128GB machine given it's roughly 2x the backbone size of V4-Flash, but it runs. ## Market context Shares of Chinese AI rivals MiniMax and Z.ai fell over 8% on the release, and Alibaba dropped more than 2%. DeepSeek is reportedly preparing for a Shanghai STAR Market listing. The pattern here is consistent: when demand goes up, DeepSeek's response is to make the model cheaper and faster rather than raise prices. ## Questions this post answers ### What is DeepSeek V4.1 Flash and how big is it? DeepSeek V4.1 Flash is a 552 billion parameter Mixture-of-Experts model released on Hugging Face, using a new Encoder-Decoder structure built with a new pre-training method and larger-scale reinforcement learning post-training. It supports native visual understanding and is the smallest model in DeepSeek's new architecture family. _Developers picking a model for multimodal or agentic workloads can track releases like this one on daily.dev._ ### Is DeepSeek V4 Pro being deprecated? Yes, DeepSeek V4 Pro is being retired and will be phased out after V4.1 Flash goes fully live. DeepSeek states that V4.1 Flash outperforms V4 Pro on performance, cost, speed, and total compute usage, making it no longer worthwhile to keep offering V4 Pro at a higher price and slower speed. _Teams relying on a pinned model id should watch for deprecations like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 5 discussions and 32 comments across x (as of 2026-09-11). **TL;DR:** Discussion centers almost entirely on the new Causal Encoder-Decoder architecture, with people debating whether it's a genuine leap deserving a V5 label and speculating on its serving/cost implications, while a few technical corrections about encoder-decoder history are traded. **Sentiment:** 45% positive · 45% mixed · 10% skeptical **The case for** - The roughly 4x smaller KV cache is seen as the real headline, more important than benchmark score gains - Some argue the architecture could make long-context ingestion and multi-session serving cheaper without slowing decode speed - The terminal-bench improvement is flagged as a standout result worth attention amid otherwise noisy benchmark gains **The pushback** - Several think calling it V4.1 undersells what is actually a major architectural pivot, calling it a marketing/naming choice - One person finds the accompanying architecture diagrams increasingly hard to follow as designs grow more complex **By community** - x (positive): Replies focus on genuine technical curiosity about the new encoder-decoder design and its serving implications, with light-hearted debate over whether it deserves a V5 name. **Hottest debate:** Whether the shift to a Causal Encoder-Decoder architecture is significant enough that the model should have been called V5 instead of V4.1. **Open questions** - Does the KV cache win come mostly from a frozen encoder prefix, or is it tied specifically to the vision/multimodal path? - How does the architecture perform in real serving conditions with many paused and resumed conversations, not just benchmarks? **Highlights** > @rasbt kv cache 3.9x smaller is the real headline, everyone chasing wrong metrics > — [0xNexqor on x](https://x.com/0xNexqor/status/2098156920649425327) > @rasbt naming it v4.1 when it's basically a new architecture is such a lab move > — [elian\_mcc on x](https://x.com/elian_mcc/status/2098253822808175053) > @rasbt cybergym and automation bench gains are close to noise, terminal bench nearly doubling everyone else is the one worth staring at > — [suqitah on x · 3 points](https://x.com/suqitah/status/2098143055366750444) > @rasbt What caught my eye is that the model got bigger, but its global KV cache got smaller. That could make long contexts cheaper to handle without making the model easier to fit on a laptop. Looks like this architecture allows handling longer prompts and also serving more users, > — [mhsamadi on x](https://x.com/mhsamadi/status/2098251989108834360) > @rasbt The interesting part is the systems implication: an encoder-decoder can make long-context ingestion cheaper without forcing every decode step to pay for it. If that holds in real workloads, agent loops—not just benchmarks—get a new cost ceiling. > — [YionxpYi on x](https://x.com/YionxpYi/status/2098257133649510884) **Source threads** - [x](https://x.com/mitsuhiko/status/2098138935507509391) · 0 points · 0 comments - [x](https://x.com/simonw/status/2098132626880262228) · 0 points · 0 comments - [x](https://x.com/scaling01/status/2098153215564853448) · 0 points · 0 comments - [x](https://x.com/rasbt/status/2098142625819672603) · 0 points · 32 comments - [x](https://x.com/antirez/status/2098123316468847044) · 0 points · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#deepseek](https://daily.dev/tags/deepseek), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DeepSeek V4.1 Flash released on Hugging Face, replacing V4 Pro","url":"https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p"},"datePublished":"2026-09-10T06:13:48.383Z","dateModified":"2026-09-11T04:03:06.391Z","description":"DeepSeek V4.1 Flash has launched on Hugging Face as a 552B parameter Mixture-of-Experts model featuring a new Encoder-Decoder architecture, a new pre-training...","image":"https://pbs.twimg.com/media/HR1ZhA3acAA7OsQ.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HR1ZhA3acAA7OsQ.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,deepseek,mixture-of-experts","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"DeepSeek V4.1 Flash released on Hugging Face, replacing V4 Pro"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/deepseek-v4-1-flash-released-on-hugging-face-replacing-v4-pro-zikp1p10p#faq","mainEntity":[{"@type":"Question","name":"What is DeepSeek V4.1 Flash and how big is it?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek V4.1 Flash is a 552 billion parameter Mixture-of-Experts model released on Hugging Face, using a new Encoder-Decoder structure built with a new pre-training method and larger-scale reinforcement learning post-training. It supports native visual understanding and is the smallest model in DeepSeek's new architecture family. Developers picking a model for multimodal or agentic workloads can track releases like this one on daily.dev."}},{"@type":"Question","name":"Is DeepSeek V4 Pro being deprecated?","acceptedAnswer":{"@type":"Answer","text":"Yes, DeepSeek V4 Pro is being retired and will be phased out after V4.1 Flash goes fully live. DeepSeek states that V4.1 Flash outperforms V4 Pro on performance, cost, speed, and total compute usage, making it no longer worthwhile to keep offering V4 Pro at a higher price and slower speed. Teams relying on a pinned model id should watch for deprecations like this via daily.dev."}}]} ```