Sep 10, 2026
DeepSeek V4: Release Date, Specs, and How to Access It (2026)
Cost Optimization
Distributed Inference
DeepSeek V4 is fully shipped: V4-Pro-0813 went GA on August 13 and its MIT weights are now on Hugging Face. Specs, new pricing, and how to access.
Both halves of DeepSeek V4 have now shipped. Here's what the GA build actually is, and which claims are verified.
The waiting is over. DeepSeek-V4-Pro-0813 left preview and went generally available in the August 12-13 rollout, live on DeepSeek's API alongside the V4-Flash model that released July 31 with open weights. The V4 generation is now fully out, and the question shifts from "when" to "what did GA actually change, and how much of the benchmark story is verified."
Here is the current picture: what's confirmed, what's still vendor-reported, and how to actually use V4 today.
Update, September 10, 2026: DeepSeek has released DeepSeek V4.1 Flash, which retires V4 Flash on the API and, from September 14, serves all V4 Pro requests at Flash pricing. The V4 details below are current for the open weights; for API pricing and the routing change, read the V4.1 Flash post.
TL;DR
- DeepSeek V4 is a family: V4-Flash (small, fast, cheap) and V4-Pro (the 1.6T flagship). Both are now released.
- V4-Flash-0731 released officially on July 31, 2026: open weights on Hugging Face under an MIT license, plus a live API
- Flash specs: 284B total parameters per the model card (the Hugging Face listing shows 304B because it counts the DSpark module) with a small activated fraction, around 13B per token as reported.
- V4-Pro went GA as build 0813 on August 13, 2026: 1.6 trillion total parameters, 49B active per token, 1M token context, up to 384K output, three reasoning modes
- The first independent read is in: Artificial Analysis scores the 0813 build 53, ranking it #3 on its leaderboard, but only three points above the much smaller Flash's 50 and short of what the model card's aggressive table implied
- DeepSeek then raised API pricing on August 17 and split it into peak and off-peak windows: Flash now runs $0.44 per million input and $1.32 per million output at peak, half that off-peak, with cache hits still priced near zero.
- Running DeepSeek in production today: DeepSeek V4 Pro and V4 Flash are live on Yotta AI Gateway, alongside V3.2 and R1.
The Release Dates, All Confirmed Now
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.
August 12-13, 2026: DeepSeek-V4-Pro-0813 left preview and went generally available on the DeepSeek API and DeepSeek Chat, plus third-party routers. The mid-August window the rumors pointed at turned out to be right.
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 model card lists 284B 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." The GA Pro's new model card now shows Pro back on top, which is what four months of post-training should buy.
Self-hosting Flash is realistic but not trivial. The full hardware requirements breakdown covers the GPU math; the short version is that 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. 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: Now Generally Available
The flagship is real now. What's confirmed about the 0813 build: 1.6 trillion total parameters with 49B active per token, a sparse mixture-of-experts design using DeepSeek's Compressed Sparse Attention and Heavily Compressed Attention, pre-trained on more than 32 trillion tokens. Context is 1M tokens with up to 384K output, and the API exposes three operating modes: non-thinking, high reasoning, and max reasoning.
The model card's benchmark table is aggressive: 80.6% on SWE-bench Verified, 90.1% on GPQA Diamond, 93.5% on LiveCodeBench, 87.5% on MMLU-Pro, 67.9% on Terminal Bench 2.0, and a 3,206 Codeforces rating.
The independent verdict started arriving within days, and it's more interesting than a simple confirmation. Artificial Analysis scored the 0813 build 53 on its Intelligence Index, ranking it #3 among all models it tracks, genuinely frontier territory. But that's only three points above V4-Flash's 50, from a model with five times the active parameters, and early third-party coverage describes the model-card table as running ahead of what independent tests show, with cybersecurity as the standout strength. The honest summary: V4-Pro is a top-three model at a tenth of frontier prices, and the card's boldest claims should still be validated on your own workload.
The GA weights arrived shortly after launch: DeepSeek-V4-Pro-0813 is on Hugging Face under MIT. The checkpoint ships alongside a DSpark variant with a built-in speculative decoding module that works with vLLM and SGLang, no separate draft model needed.
What DeepSeek V4 Costs
DeepSeek published the API rate card, and the headline is the cache economics:
| Per 1M tokens | V4-Flash | V4-Pro |
| Input, cache miss | $0.44 peak / $0.22 off-peak | $1.32 peak / $0.66 off-peak |
| Input, cache hit | $0.014 peak / $0.007 off-peak | $0.044 peak / $0.022 off-peak |
| Output | $1.32 peak / $0.66 off-peak | $3.96 peak / $1.98 off-peak |
Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. Everything else is off-peak at half the peak rate.
Cache hits are priced at roughly a 97 percent discount, which changes the math for agentic and long-context workloads where most input tokens repeat. The peak and off-peak split took effect August 17, so time-shift batch work to off-peak hours and the whole rate card halves.
For context: V4-Pro's $3.96 peak output ($1.98 off-peak) sits against $6 for Qwen 3.8-Max, $15 for Kimi K3, and $30 for GPT-5.6 Sol. These are different capability classes, but the independent numbers back the value story: a top-three model at a tenth of frontier-class prices. 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: V4-Pro serves in GA (concurrency limit 500), Flash in public beta (limit 2,500). Both speak OpenAI ChatCompletions, Anthropic Messages, and DeepSeek's own Responses format, so existing client code ports with a base URL change.
The weights: DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813 are both 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 and choosing an inference engine cover the stack decisions.
The production route: DeepSeek V4 Pro and V4 Flash are live on Yotta AI Gateway behind one OpenAI-compatible API, alongside V3.2, R1, Qwen, GLM, Kimi, and the rest of the catalog, with Claude being added.
Frequently Asked Questions
When is the DeepSeek V4 release date?
In stages, all now complete: family preview April 24, 2026; V4-Flash with open weights July 31; V4-Pro GA (build 0813) in the August 12-13 rollout.
Is DeepSeek V4 open source?
Yes, both halves now. V4-Flash-0731 and V4-Pro-0813 are both on Hugging Face under MIT licenses. Flash is the one most teams can realistically run; Pro's checkpoint is far larger.
What does DeepSeek V4 cost?
Current API pricing after the August 17 increase: Flash at $0.44 per million input tokens and $1.32 output at peak, Pro at $1.32 in and $3.96 out at peak. Off-peak hours run half price, and cache hits are discounted roughly 97 percent.
Can I self-host DeepSeek V4?
Flash, yes: 166.9 GB of MIT-licensed weights, multi-GPU territory. Pro, now also yes on paper: DeepSeek published the 0813 weights on Hugging Face under MIT. In practice a 1.6 trillion parameter checkpoint is multi-node cluster territory, so most teams will use the API or a hosted provider for Pro and keep self-hosting for Flash.
Is DeepSeek V4 better than Qwen 3.8?
The evidence gap widened: V4-Pro now has both a published model card and independent scores (53 on Artificial Analysis, ranked #3), while Qwen 3.8-Max still has neither. Head-to-head testing on your own workload remains the real answer, but DeepSeek is currently the evidence-backed option of the two.
Is DeepSeek V4 on Yotta?
Yes. DeepSeek V4 Pro and V4 Flash are both live on Yotta AI Gateway, alongside DeepSeek V3.2 and DeepSeek R1, behind one OpenAI-compatible API with flat pricing.
Bottom Line
DeepSeek finished shipping V4: open weights for both tiers, and a GA flagship. The first independent answer arrived fast: a 53 on Artificial Analysis and a #3 ranking, which confirms frontier-class capability while trimming the model card's boldest claims down to size.
The practical move: evaluate V4 through the API now, keep your endpoint OpenAI-compatible, and let the independent numbers decide how much traffic it earns. If you want the DeepSeek line in production today, DeepSeek V4 Pro, V4 Flash, V3.2, and R1 are live on Yotta AI Gateway, one API key away.
