--- title: "DeepSeek V4: Release Date, Specs, and How to Access It (2026)" slug: deepseek-v4-release-date-specs-how-to-access-2026 description: "Half of DeepSeek V4 already shipped: Flash is live with open weights and a public API. The full V4-Pro is the release date everyone is asking about. Here's what's confirmed." author: "Yotta Labs" date: 2026-08-07 categories: ["Inference"] canonical: https://www.yottalabs.ai/post/deepseek-v4-release-date-specs-how-to-access-2026 --- # DeepSeek V4: Release Date, Specs, and How to Access It (2026) ![](https://cdn.sanity.io/images/wy75wyma/production/df66dbedb1be2f683029257aeaa9a660e07d84a4-2240x1260.png) If you're trying to pin down "the DeepSeek V4 release date," the confusing answer is that part of V4 already shipped. DeepSeek-V4-Flash went official on July 31, 2026 with open weights on Hugging Face and a live API. The flagship, DeepSeek-V4-Pro, is still in preview, and its general availability is the date nobody has confirmed. That split matters because the two models are built for different jobs, priced very differently, and sitting at opposite points of the release cycle. Here is what's confirmed, what's reported, and how to actually use V4 today. ## TL;DR - DeepSeek V4 is a family: V4-Flash (small, fast, cheap) and V4-Pro (the 1.6T flagship). Both previewed April 24, 2026 - V4-Flash-0731 released officially on July 31, 2026: open weights on Hugging Face under an MIT license, plus a live API - Flash specs: roughly 300B total parameters (the Hugging Face repo lists 304B) with a small activated fraction, around 13B per token as reported. 1M token context and up to 384K output are reported alongside the release - V4-Pro remains preview-only: 1.6 trillion total parameters, 49B active, per the April preview. DeepSeek's official line on full release timing is "soon" - Third-party reports point to a mid-August 2026 GA window for the full V4. DeepSeek has not confirmed a date - Published Flash API pricing: $0.14 per million input tokens, $0.28 per million output, with cache hits priced near zero - Running DeepSeek in production today: DeepSeek V3.2 and DeepSeek R1 are live on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) ## The Release Date Question, Answered Honestly Three dates matter, and only two of them are confirmed. April 24, 2026: DeepSeek previewed the V4 family, including the V4-Pro preview (1.6T parameters, 49B active) and early Flash builds, under MIT licensing. Confirmed. July 31, 2026: DeepSeek-V4-Flash-0731 released officially. Open weights live on Hugging Face (166.9 GB across 48 shards), API serving in public beta. Confirmed. Mid-August 2026: the reported window for full V4 / V4-Pro general availability. This comes from third-party coverage, not DeepSeek, and earlier reports had said August 10 to 20 before slipping to mid-month. Treat it as a rumor with reasonable smoke behind it. DeepSeek's own public position is still just "soon." We will update this post when the Pro date firms up either way. ## DeepSeek V4-Flash: The Half That Already Shipped Flash is the small-and-fast line, and the 0731 build is a real release, not a preview: weights you can download today under MIT, one of the most permissive licenses in the frontier field. The Hugging Face repo lists 304B total parameters, with a far smaller activated count per token, reported around 13B. That shape is the point: a sparse model that serves cheap and fast while keeping frontier-adjacent capability. Reported context is 1M tokens with up to 384K output, and the API exposes low, high, and max reasoning effort levels. One result worth pausing on: DeepSeek's own model card shows Flash-0731 outscoring the V4-Pro preview across its published agentic benchmarks. Artificial Analysis independently scored the 0731 build 10 points above April's Flash on its Intelligence Index. Read that less as "Flash beats Pro" and more as "the Flash line improved fast enough to lap a four-month-old preview." It also means the Pro that eventually ships GA will need to clear a bar its own little sibling already set. Self-hosting Flash is realistic but not trivial: 166.9 GB of weights puts you in multi-GPU territory before KV cache, the middle ground between a single-GPU 27B and [Kimi K3's 1.56 TB cluster-scale deployment](https://www.yottalabs.ai/post/kimi-k3-hardware-requirements-gpu-memory-2026). The same rule applies as with every sparse MoE model: low active parameters save you compute, not memory. All 300B-ish parameters need to live in VRAM regardless of how few fire per token. ## DeepSeek V4-Pro: The One Everyone Is Waiting For The flagship. What the April preview established: 1.6 trillion total parameters, 49B active per token, MIT licensing intent, and preview API access with Responses API and coding-tool support expanding through early August. What doesn't exist yet: a GA date, final benchmarks for the release build, or open weights for the Pro tier. Until those land, V4-Pro is a strong preview with the same evidence gap every unreleased flagship has, and anyone quoting a firm release date is passing along a rumor. If the mid-August window holds, this becomes the next major open-lab release cycle within days. That is worth planning around even before it's confirmed. ## What DeepSeek V4 Costs DeepSeek published the API rate card, and the headline is the cache economics: Cache hits are priced at roughly a 98 percent discount, which changes the math for agentic and long-context workloads where most input tokens repeat. A peak-hour surcharge has been announced but is not yet in effect, so budget for list price doubling at peak before you commit to a cost model. For context against the rest of the field: Flash's output price is $0.28 against $6 for Qwen 3.8-Max and $15 for Kimi K3. These are different capability classes, so cost per token is not cost per solved task. Validate on your own workload before building a budget on any vendor's rate card. ## How to Access DeepSeek V4 Today Three real options right now. The API: Flash-0731 serves as `deepseek-v4-flash` in public beta, with Pro available in preview. This is the fastest way to evaluate either. The weights: DeepSeek-V4-Flash-0731 is on Hugging Face under MIT. If you have multi-GPU capacity, you can serve it yourself; our guides to [the best open-source LLMs of 2026](https://www.yottalabs.ai/post/best-open-source-llms-2026) and [choosing an inference engine](https://www.yottalabs.ai/post/vllm-vs-sglang-which-inference-engine-should-you-use-in-2026) cover the stack decisions. The production route: DeepSeek V3.2 and DeepSeek R1 are live on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) today behind one OpenAI-compatible API, alongside Claude, Qwen, GLM, and the rest of the catalog. V4 is not on the Gateway catalog yet. If it lands there once GA arrives, switching from V3.2 is a config change, not a migration, which is the cleanest way to be ready for a model that doesn't have a confirmed date. ## Frequently Asked Questions **When is the DeepSeek V4 release date?** V4-Flash already released: July 31, 2026, with open weights and a public API. The full V4-Pro GA is unconfirmed; third-party reports point to mid-August 2026, and DeepSeek officially says "soon." **Is DeepSeek V4 open source?** V4-Flash-0731's weights are on Hugging Face under an MIT license, which is genuinely permissive. V4-Pro has no public weights yet. As with most frontier releases, training data and code are not open, so "open weights" is the accurate label. **What does DeepSeek V4 cost?** Published API pricing: Flash at $0.14 per million input tokens and $0.28 output; Pro preview at $0.435 in and $0.87 out. Cache hits are discounted roughly 98 percent, and an announced peak-hour surcharge is not yet active. **Can I self-host DeepSeek V4?** Flash, yes: 166.9 GB of MIT-licensed weights, which means multi-GPU capacity before KV cache. Pro, no: weights have not been released. **Is DeepSeek V4 better than Qwen 3.8?** No verified head-to-head exists. The evidence situations differ: Flash-0731 has third-party index scores, while Qwen 3.8-Max still has no published benchmark table at all. Our [Qwen 3.8 benchmarks tracker](https://www.yottalabs.ai/post/qwen-3-8-benchmarks-what-is-verified-2026) covers that side of the ledger. **Is DeepSeek V4 on Yotta?** Not yet. DeepSeek V3.2 and DeepSeek R1 are live on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway). This post will note it if V4 joins the catalog. ## Bottom Line DeepSeek V4 is half-shipped. Flash is real today: open weights, MIT license, published pricing, and third-party scores, which is more evidence than most 2026 flagship launches managed on day 30. The Pro flagship is the open question, with a reported mid-August window and nothing confirmed. The practical move: evaluate Flash now, through the API or the weights, and have your integration ready so Pro GA is an endpoint swap rather than a project. If you want the DeepSeek line in production this week, [DeepSeek V3.2 and R1 on Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) are the live options, one API key away.