DeepSeek V4.1 Pro: The Flagship Tier DeepSeek Names But Has Not Shipped

DeepSeek V4.1 Pro is the Pro tier of the V4.1 line — the larger, flagship-class sibling of DeepSeek V4.1 Flash, the 552-billion-parameter multimodal model DeepSeek released on September 10, 2026. Flash is the tier built for input-heavy agent work: DeepSeek's own release note calls it "the smallest model in our new architecture family," engineered so that long contexts stay cheap. The Pro tier is the one aimed at the hardest reasoning and agentic tasks, where the family's biggest backbone is supposed to buy accuracy. The distinction that matters for anyone reading this today is narrower than the marketing either side of the family attracts: DeepSeek V4.1 Flash can be called right now, and DeepSeek V4.1 Pro cannot be called anywhere at all.

This page is the reference for that second name — what it is, what it would be built on, and what is actually established about it — checked against DeepSeek's own documentation rather than against the rumour cycle that has circled it since September. It is not a launch piece, because nothing has launched: DeepSeek has named DeepSeek V4.1 Pro exactly once in public, in its own September 10 release note, and has published nothing about it since.

What the Pro tier is for, and why it has no card yet

In the V4 generation, the tier split is easy to state because both sides shipped. DeepSeek V4 Flash is the efficiency model — a 284-billion-parameter mixture-of-experts with 13 billion parameters active — while DeepSeek V4 Pro is the flagship: a 1.6-trillion-parameter mixture-of-experts with 49 billion active. Both accept a 1M-token context and return up to 384K tokens of output. Pro is roughly four and a half times the price of Flash per token and carries a fifth of the concurrency allowance, and it is text-only: on DeepSeek's current price sheet the vision row reads "Not supported" for Pro and a check mark for Flash.

The V4.1 line restates that split rather than inventing a new one. The model that shipped, DeepSeek V4.1 Flash, is a 552B-parameter mixture-of-experts on a new Causal Encoder-Decoder architecture that activates 8B parameters during prefill and 16B during decode, natively reads images, and serves the same 1M context and 384K output. DeepSeek describes it as the smallest member of the new architecture family — a sentence that only makes sense if there are larger members planned. DeepSeek V4.1 Pro is the name it has used for the biggest of them.

What exists for the Pro name today, as of October 13, 2026, is one sentence. In the September 10 release note that shipped DeepSeek V4.1 Flash, DeepSeek wrote that from 04:00 UTC on September 14 all requests to deepseek-v4-pro would route to V4.1 Flash and be billed at Flash rates, and that this would continue "until V4.1-Pro launches." That plan was withdrawn after developers objected to a production model being swapped underneath them — the same changelog entry records that DeepSeek decided instead to keep serving DeepSeek V4 Pro after September 14 with billing unchanged. The routing plan died. The name survived it, and no release date has been announced for it since.

The specification, and whose model card it comes from

DeepSeek V4.1 Pro has no model card. DeepSeek's changelog has no entry for it, the company's price sheet has no row for it, and nothing named for it exists in the deepseek-ai organisation on Hugging Face, whose newest repository is still DeepSeek-V4.1-Flash from September 10. So there is no context length, no parameter count, no modality statement and no licence term that can be attributed to DeepSeek V4.1 Pro itself. Anyone who publishes those figures as if they were the model's is quoting the family and calling it the model.

What can be stated with a source is the Pro tier of the previous generation, which is the closest thing DeepSeek has documented, and the shipped V4.1 model the Pro tier would stand above. Each figure below is attributed to the document it came from.

• Backbone and active parameters — DeepSeek V4 Pro is 1.6T total with 49B active; DeepSeek V4.1 Flash is 552B with 8B active at prefill and 16B at decode. Both per DeepSeek's Hugging Face model cards.

• Context and output — 1M tokens of context and a 384K maximum output, listed identically for both tiers on DeepSeek's Models & Pricing page. No separate figure exists for DeepSeek V4.1 Pro.

• Modality — DeepSeek V4.1 Flash takes text and images and returns text; its model card is tagged image-text-to-text, and DeepSeek's price sheet marks vision as supported. The Pro column on that same sheet marks vision "Not supported." Whether a V4.1 Pro would close that gap is unpublished.

• Reasoning behaviour — thinking is the default mode on both served models, with non-thinking available, and DeepSeek's changelog says the V4 Pro and V4 Flash tiers support three thinking-effort levels: low, high and max. DeepSeek's own guidance is to use low for simple tasks, high for daily agent work and max for harder scenarios.

• Licence — MIT, for the entire V4 and V4.1 line that has shipped: DeepSeek-V4-Pro, DeepSeek-V4-Flash, DeepSeek-V4-Flash-0731 and DeepSeek-V4.1-Flash all carry `license:mit` on their Hugging Face cards. Nothing about DeepSeek V4.1 Pro's terms has been published.

• Benchmarks — none exist for DeepSeek V4.1 Pro. DeepSeek's V4.1 technology report, a 51-page document published alongside Flash under the title "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," is a Flash paper throughout: the strings "V4.1-Pro" and "V4.1 Pro" appear nowhere in it, and its comparison tables use DeepSeek-V4-Pro as the baseline the new model beats, not as a tier it belongs to.

Where it sits in the V4.1 line and in the wider V4 family

DeepSeek's shipped releases run in a clear order. The V4 generation arrived on April 24, 2026, exposed through the API as deepseek-v4-pro and deepseek-v4-flash. The Flash tier went official on July 31 with a re-post-trained checkpoint; the Pro tier followed to general availability on August 13 as DeepSeek-V4-Pro-0813, superseding the April preview; an experimental vision model, DeepSeek-V4-Flash-Vision-Exp, landed on August 21 and was retired within three weeks, absorbed into the V4.1 release. Then DeepSeek V4.1 Flash on September 10, which is where the family stands today. DeepSeek V4.1 Pro would be the first V4.1-tier model above Flash, and in the V4 generation that tier trailed the efficiency model by about a fortnight. The Pro name has now trailed DeepSeek V4.1 Flash by more than a month, which is the one measurable thing about its timing and a fair reason to stop treating it as imminent.

How a reader reaches it: by neither route

Access is the shortest section on this page because there is nothing to access. Per DeepSeek's own API reference, the models the platform currently serves are exactly two: deepseek-flash, which is DeepSeek-V4.1-Flash, and deepseek-v4-pro, which is DeepSeek-V4-Pro-0813. There is no DeepSeek V4.1 Pro endpoint, no rate limit published for one, and no name to pass in a request. Weights are the same story: no repository, no checkpoint, nothing to download in any format, and therefore no GGUF, MLX or NVFP4 conversion of a model that does not exist. The only priced comparison a reader can run today is between the two tiers that did ship — DeepSeek V4.1 Flash at $0.15 input and $0.60 output per million tokens off-peak, doubling to $0.30 and $1.20 in the weekday peak windows, against DeepSeek V4 Pro at $0.66 and $1.98 off-peak, doubling to $1.32 and $3.96. Those are DeepSeek's own numbers on its own Models & Pricing page, with peak hours defined as 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. No successor price has been hinted at.

The speculative pages already exist

Two pages on this site already track the guessing, and readers who want that chronology should use them rather than this page: the V4.1 Pro leak timeline covers what was reported and when, and the September release-window watch covers the window that was expected to close at the end of that month. This page deliberately does not repeat either of them. What it does instead is state the factual position of a model that has a name, a family and a described role, and no published specification.

What is not settled

Almost everything, and it is worth being specific about which parts are open rather than vague. Two things are genuinely unknown: whether DeepSeek V4.1 Pro exists as a trained checkpoint at all, and whether it will inherit the Causal Encoder-Decoder architecture, the native multimodal input and the MIT licence that the rest of the V4.1 line carries — none of which is a safe assumption from Flash, because the V4 generation ran the two tiers on deliberately different shapes, including Pro being the text-only one. One thing is worth watching rather than assuming: DeepSeek says Flash comprehensively surpasses DeepSeek V4 Pro on performance, cost, speed and total runtime, which is the company's own claim about its own flagship, and it leaves the Pro tier's purpose in the V4.1 line redefined — larger in name, cheaper in practice, or both.

What a reader can call today

OrcaRouter does not serve DeepSeek V4.1 Pro, and no amount of adjacent availability makes that true: there is no model id matching deepseek-v4.1-pro on any spelling in our catalogue, and none in the vendor's either. What we do serve is the tier that shipped. The V4.1 model on our list is DeepSeek V4.1 Flash, at $0.15 input and $0.60 output per million tokens — DeepSeek's off-peak list price passed through with no markup — with 1M tokens of context, 384K of maximum output, text and image input, tool calls, JSON output and the Responses and Anthropic-compatible endpoints. The flagship of the previous generation is also served, and it is a different model with a different name: DeepSeek V4 Pro, at $0.66 and $1.98, which is the text-only 1.6T tier described above and not a V4.1 model.

That substitution is worth stating plainly rather than glossing, because the two served models answer different questions. If the reason you searched for DeepSeek V4.1 Pro is that you want the strongest DeepSeek model you can put a production path behind, DeepSeek V4 Pro is the honest answer and it is one request away. If what you actually want is the new architecture, the native image input and the reduced KV-cache cost, DeepSeek V4.1 Flash is the only released model that has any of those, and its price per token is a quarter of Pro's on input and a third on output. Running a slice of traffic against DeepSeek V4.1 Flash alongside your existing model is a one-key change here, and automatic failover means a call that breaks on one provider path does not take the workflow with it — which is the right posture for a family whose tier above Flash has no date on it.

How to tell when this page is out of date

Three things would falsify it, and all three are checkable in one page load each. A changelog entry on DeepSeek's API documentation naming DeepSeek V4.1 Pro, since that is where every V4 release has been announced and its newest entry today is September 10. A new repository in the deepseek-ai organisation with V4.1-Pro in the name; the organisation's newest repository today is DeepSeek-V4.1-Flash, created September 10, 2026. Or a third column on DeepSeek's Models & Pricing page, which today lists exactly two models and is where the vendor states context, modality, features, concurrency and per-token rates. Until one of those three moves, the position is simple: the tier exists as a name DeepSeek used and has not used since, the smaller model is shipping and priced, and the flagship of the previous generation is the closest thing to a top-tier DeepSeek model anyone can call.