Last updated: October 2026
Key Takeaways
- DeepSeek V4-Pro is no longer a preview. The official V4-Pro-0813 shipped on August 13, 2026 with MIT-licensed open weights, and it is the model behind the deepseek-v4-pro API name today.
- API prices went up, not down. Since August 16, DeepSeek bills by time of day: V4-Pro costs $0.66 per million input tokens (cache miss) and $1.98 per million output tokens off-peak, and double that at peak.
- The V4-Flash API names now answer with V4.1-Flash, a different and larger model released September 10 and updated to the 1017 checkpoint on October 17. DeepSeek's API no longer serves V4-Flash-0731; the open weights on your own disk still do.
Yes, DeepSeek V4 is now fully released. As of October 18, 2026, V4-Pro has left preview as V4-Pro-0813, an MIT-licensed open-weight model of about 1.65 trillion parameters that costs $0.66 input and $1.98 output per million tokens off-peak. V4-Flash-0731 is the final V4-Flash, and since September 10 its API names route to the newer V4.1-Flash, which became V4.1-Flash-1017 on October 17.
This page is a status ledger: every claim below carries a date and a source.
What Changed Since April: The Ledger
Eleven events have changed what DeepSeek V4 is since launch day; this table is the catch-up.
| Date | What changed | Source |
|---|---|---|
| April 23, 2026 | V4-Pro and V4-Flash launch as an open-weight preview under the MIT License, with V4-Pro output priced at $3.48 per million tokens | DeepSeek announcement |
| May 31, 2026 | The promotional 75% V4-Pro price cut becomes permanent: $0.435 input (cache miss) and $0.87 output per million tokens | DeepSeek pricing docs; InfoWorld |
| July 24, 2026 | The legacy deepseek-chat and deepseek-reasoner names are retired at 15:59 UTC; requests to them now fail | DeepSeek API docs |
| July 31, 2026 | V4-Flash-0731 ships in public beta under the existing API name, with open weights the same day | DeepSeek change log; Hugging Face |
| August 13, 2026 | V4-Pro-0813 is released as the official V4-Pro on app, web, and API, with MIT open weights and three reasoning-effort levels | DeepSeek change log; Hugging Face |
| August 16, 2026 | Peak and off-peak pricing takes effect at 16:00 UTC; standard rates rise, and off-peak rates are half of peak | DeepSeek announcement; InfoWorld |
| August 21, 2026 | The experimental V4-Flash-Vision-Exp opens on the API | DeepSeek change log |
| September 10, 2026 | V4.1-Flash is released with open weights and a new API name, deepseek-flash; V4-Flash and V4-Flash-Vision-Exp are retired, and their names route to V4.1-Flash | DeepSeek change log |
| September 14, 2026 | A planned switch of every deepseek-v4-pro call to V4.1-Flash is called off; V4-Pro keeps running at unchanged billing | DeepSeek change log |
| October 17, 2026 | V4.1-Flash-1017 replaces V4.1-Flash under the same deepseek-flash name and prices, MIT weights the same day; the Search and Fetch APIs open to all API users | DeepSeek change log; Hugging Face |
| October 20, 2026 | DeepSeek open-sources DeepInfer, its production inference engine, under Apache 2.0, plus a deepseek-recipe library that matches its API's request handling; V4.1-Flash-1017 only, on NVIDIA SM100 data-center GPUs and Huawei Ascend 950 NPUs, with no cache or load-balancing layer | DeepSeek on X; GitHub; The New Stack |
Rows verified against DeepSeek's change log, its pricing page, and Hugging Face on October 18, 2026; the October 20 row against the DeepInfer repository.
Is DeepSeek V4 Fully Released Now?
Yes. On August 13, 2026, DeepSeek's release announcement made V4-Pro generally available on its app, web, and API, and the V4-Pro-0813 model card calls it "the official release of DeepSeek-V4-Pro, superseding the preview version." The weights went up on Hugging Face the same day under the MIT License.
Our August call that V4-Pro was still a preview held: the July 20 "general availability" headlines traced to grayscale testing, and the real release came 24 days later. DeepSeek also open-sourced its Harness agent tool (version 0.1, MIT) that week, per DataNorth's report.
V4-Pro-0813 keeps the preview's architecture and attaches a DSpark speculative-decoding module, which raises the stored parameter count from about 1.60 trillion to 1.65 trillion and the download to about 893GB. Reasoning effort is now a setting with three levels, low, high, and max.
V4-Flash-0731 remains the final V4-Flash checkpoint, but since September 10 DeepSeek's API no longer serves it: the deepseek-v4-flash name routes to V4.1-Flash, a new architecture (1017 checkpoint since October 17) with its own hardware reality check.
Verified Specifications: V4-Pro-0813, V4-Flash-0731, and V4.1-Flash
All three are Mixture-of-Experts models: a large total parameter count with a small fraction active on any given token, all under the MIT License.
| Specification | V4-Pro-0813 | V4-Flash-0731 | V4.1-Flash (1017) |
|---|---|---|---|
| Released | August 13, 2026 (official) | July 31, 2026 (final V4-Flash) | September 10, 2026; 1017 update October 17 |
| Total parameters | 1.6 trillion, plus a DSpark module (1.65 trillion stored) | 284 billion, plus a DSpark module (304 billion stored) | 552 billion backbone plus 196 billion of Engram memory |
| Active parameters per token | 49 billion | 13 billion | 8 billion reading input, 16 billion generating |
| Download size | About 893GB | 166.9GB | 510.30GB |
| Context window | 1 million tokens | 1 million tokens | 1 million tokens |
| Inputs | Text | Text | Text and images |
| DeepSeek API name today | deepseek-v4-pro | None; its old names route to V4.1-Flash | deepseek-flash |
From DeepSeek's model cards, Hugging Face file listings, and API pricing page, checked October 18, 2026. At its April launch, V4-Pro overtook Kimi K2.6 (1.1 trillion) as the largest open-weight model released, per TechCrunch's release-day reporting.
Why V4 Is Architecturally Interesting
The consequential part of V4 is not its parameter count but how much compute and memory DeepSeek removed from long-context inference. In DeepSeek's published figures, at a 1 million token context V4-Pro uses roughly 27% of the single-token inference compute and 10% of the KV cache memory of its V3.2 predecessor; V4-Flash pushes that to roughly 10% and 7%.
Four choices drive the gains. A hybrid attention design pairs Compressed Sparse Attention with Heavily Compressed Attention, so long contexts are processed without comparing every token to every other token at full precision. Manifold-Constrained Hyper-Connections stabilize signal propagation at extreme depth. The Muon optimizer handles training stability. And quantization-aware training in mixed FP4 and FP8 precision keeps the downloads in the hundreds of gigabytes rather than multiple terabytes. V4.1-Flash adds a fifth idea, Engram: 196 billion parameters of lookup memory that the model reads rather than computes, which home setups keep on an SSD.
One supply-chain note has aged well: DeepSeek validated V4's expert-parallel scheme on both Nvidia GPUs and Huawei Ascend NPUs, continuing the pattern we covered when Z.ai trained GLM-5.1 entirely on Huawei silicon. The Chinese open-weight frontier increasingly does not depend on Nvidia hardware, which matters for export-control watchers.
What DeepSeek V4 Costs Now
Prices went up. Effective 16:00 UTC on August 16, DeepSeek moved to time-of-day billing and raised its rates; InfoWorld calculated V4-Pro output increases of 127 to 355 percent, depending on the hour, and DeepSeek cited demand straining its capacity. The V4.1-Flash launch on September 10 brought the Flash tier back down, and the 1017 update on October 17 left prices unchanged while opening the Search and Fetch APIs to all users at $0.75 per 1,000 successful calls from November 1, with no data-retention statement. The table shows DeepSeek's off-peak rates from its pricing page on October 18, 2026, beside the closed-frontier list prices in effect the same day.
| Model | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| DeepSeek V4.1-Flash-1017 (deepseek-flash), off-peak | $0.003 | $0.15 | $0.60 |
| DeepSeek V4-Pro-0813 (deepseek-v4-pro), off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek V4-Pro-0813, peak hours | $0.044 | $1.32 | $3.96 |
| Gemini 3.1 Pro Preview | See note | $2.00 | $12.00 |
| Claude Opus 5.5 | See note | $4.00 | $20.00 |
| GPT-6 Astra | See note | $10.00 | $50.00 |
| Claude Fable 5.1 | See note | $10.00 | $50.00 |
USD per million tokens, list prices as of October 18, 2026, from each vendor's own pricing page. DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; peak rates are double. The Gemini rate applies to prompts up to 200K tokens and the GPT-6 Astra rate to requests up to 272K input tokens. Competitors price caching under different schemes, so cache-hit columns are shown only where comparable, and tokenizers differ, so per-token comparisons are approximate.
The arithmetic on output pricing: off-peak V4-Pro at $1.98 is about one-tenth the price of Claude Opus 5.5 and one twenty-fifth of GPT-6 Astra or Claude Fable 5.1. At peak, V4-Pro costs about one-fifth of Opus 5.5. V4.1-Flash at $0.60 still sits below one-eightieth of the Fable 5.1 rate.
Two practical caveats. Thinking-mode reasoning tokens bill at the output rate, so reasoning-heavy calls cost more than the headline implies. And DeepSeek's peak window falls overnight for North American users, 9 p.m. to midnight and 2 a.m. to 6 a.m. Eastern Daylight Time, so batch jobs scheduled for those hours pay double. This market has repriced itself three times this year; check the rate card before committing a production budget.
Benchmarks: What Is Verified and What Is Not
DeepSeek's technical report describes V4 as trailing the closed frontier by roughly three to six months on the hardest cross-domain reasoning, an unusually honest self-assessment. At launch, V4-Pro's 80.6% on SWE-bench Verified was the strongest open-weight score on real-world software engineering.
The newer figures come with the same asterisk as before. DeepSeek reports 87.9 on Terminal-Bench 2.1 for V4-Pro-0813 and 90.6 for V4.1-Flash, which means its newest small model out-scores its flagship on DeepSeek's own table. Those scores come from DeepSeek's own runs, and independent replication is still thin; our breakdown of the V4-Flash-0731 release reads those asterisks in detail.
The Data Sovereignty Question
Using DeepSeek's hosted API means your prompts and outputs traverse Chinese infrastructure. For most personal use, that is not categorically different from trusting a US operator's servers: you are handing tokens to someone else either way. For regulated industries, IP-sensitive work, or data-residency obligations, it is a harder question, and DeepSeek's history adds context: Anthropic's February 2026 distillation report named it among the Chinese labs accused of extracting Claude's capabilities through proxy accounts.
On July 31 the deepseek-v4-flash name moved to a new checkpoint overnight. On September 10 the same name began answering with a different architecture, V4.1-Flash, and on October 17 it was swapped for the 1017 checkpoint. Every deepseek-v4-pro call was scheduled to follow on September 14, until user demand changed DeepSeek's plan. An API model's identity belongs to its operator; weights on your own disk change only when you change them. The mitigations are unchanged: self-host so the model cannot change without you, use a non-DeepSeek host after reading its data terms, or send only non-sensitive work to the API. The MIT License keeps the first option real, and since October 20 the serving engine is open too, so other hosts can build on DeepSeek's own serving code.
Can You Run V4 at Home?
V4-Pro-0813, no: at about 893GB it is a multi-GPU data-center deployment, an open-weight enterprise model rather than a local one. V4-Flash-0731, yes: mainline llama.cpp runs it, and the recommended 3-bit build fits 128GB unified-memory machines, as our V4-Flash hardware reality check details. V4.1-Flash runs today on two linked 128GB DGX Sparks through community recipes, with its Engram tables on an SSD; our V4.1-Flash reality check maps every tier; DeepSeek's new DeepInfer engine targets data-center GPUs only. Below 128GB, run a model that fits well instead; our guides to the best local AI models by VRAM and memory and the best hardware for local AI map the tiers.
Not sure your own machine can run DeepSeek V4-Flash, the V4 model a home machine can realistically hold? Our LLM VRAM calculator grades it (and V4-Pro, from its model picker) against your graphics card, Mac or mini PC, names the file to download and prints the command to start it. It runs in your browser, with no sign-up.
Check your hardware for DeepSeek V4-Flash
Frequently Asked Questions
Is DeepSeek V4 fully released or still a preview?
Fully released. V4-Pro left preview on August 13, 2026 as V4-Pro-0813, with MIT-licensed open weights on Hugging Face and general availability on DeepSeek's app, web, and API. V4-Flash-0731, released July 31, is the final V4-Flash; on DeepSeek's API it has been replaced by V4.1-Flash since September 10, 2026, updated to 1017 on October 17.
What does the DeepSeek API cost right now?
As of October 18, 2026, off-peak, deepseek-v4-pro costs $0.66 per million input tokens on a cache miss, $0.022 on a cache hit, and $1.98 per million output tokens. deepseek-flash costs $0.15, $0.003, and $0.60. Rates double during peak hours, 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Prices rose on August 16, and the Flash tier came back down on September 10.
What is the difference between V4-Pro, V4-Flash, and V4.1-Flash?
V4-Pro-0813 is the flagship: 1.6 trillion parameters with 49 billion active per token, text only. V4-Flash-0731 is the efficient V4 sibling: 284 billion total and 13 billion active, and it runs on a 128GB machine. V4.1-Flash (1017 since October 17) is a newer architecture: a 552 billion parameter backbone plus 196 billion of Engram memory, with image input. All three are MIT-licensed.
Is DeepSeek shutting down the V4-Pro API?
No. DeepSeek announced on September 10 that deepseek-v4-pro requests would route to V4.1-Flash from September 14, 2026, then reversed course and kept serving V4-Pro "in response to user demand," with billing unchanged. Its pricing page lists the deepseek-v4-pro name as V4-Pro-0813, without image input.
What happened to deepseek-v4-flash, deepseek-chat, and deepseek-reasoner?
deepseek-chat and deepseek-reasoner were retired on July 24, 2026, and requests to them fail. deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but route to V4.1-Flash, and since October 17 to its 1017 checkpoint; DeepSeek asks integrations to switch to deepseek-flash. The name is the same, but the model answering is no longer V4-Flash.
Can I run DeepSeek V4 on my own hardware?
V4-Pro-0813 needs data-center hardware for its 893GB of weights. V4-Flash-0731 runs locally through mainline llama.cpp with a memory floor around 110GB, which means a 128GB unified-memory machine. V4.1-Flash runs on two linked DGX Sparks through community recipes. Below 128GB, smaller open models such as Gemma 4 and Qwen variants are the practical choices.
Is it safe to send sensitive data to DeepSeek's API?
It depends on your threat model. Prompts to the hosted API traverse Chinese infrastructure, which for regulated or IP-sensitive work is a genuine data-residency concern, compounded by the fact that the model behind an API name can change without notice, as it did on July 31, September 10 and October 17. Self-hosting an open-weight V4 model keeps your data on your network.
