# DeepSeek V4: Open Weights Reach the Frontier

By Amir Teymoori

---

My verdict up front. DeepSeek V4-Flash is the best cost-per-useful-token I can currently get for coding agents, bulk extraction, and anything where I'm burning tokens by the billion. V4-Pro is a genuinely frontier-adjacent model that almost nobody reading this will self-host. Both are MIT licensed. Only one of them is realistically runnable outside a datacenter.

That split matters more than the benchmark charts.

## What Actually Shipped

DeepSeek V4 went GA on July 20, 2026, moving out of preview with noticeably stronger agentic behavior, math reasoning, and code generation. The flagship is V4-Pro. The cheaper sibling, V4-Flash, got a refresh on July 31 as `DeepSeek-V4-Flash-0731`, which per DeepSeek's own notes is the same architecture and size, re-post-trained for agent and coding gains. Same API name, same price, different model behind the endpoint.

Worth flagging if you pin behavior: the model answering your `deepseek-v4-flash` calls changed on July 31 without a version bump in the string. I re-ran my eval set. Results held up.

DeepSeek isn't operating alone here. Qwen 3.7 Flash landed July 27 from Alibaba, and Moonshot shipped Kimi K3 on July 16. Three serious open-weight releases inside sixteen days, all from Chinese labs.

## The Architecture Headline

V4-Pro is a sparse mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token, built around a native 1-million-token context window with output up to roughly 384K tokens. V4-Flash is the same idea at a quarter of the scale: 284B total, 13B active, same 1M context.

Sparse MoE is the whole trick: knowledge capacity of a 1.6T model, compute cost of 49B parameters per forward pass. That's why the API pricing is what it is.

The context window is native rather than bolted on through position extrapolation, and in my testing it degrades more gracefully past 200K than the retrofitted long-context models I used last year. Still, I wouldn't stuff 800K tokens into a prompt and trust the middle of it. Retrieval discipline beats brute force, and if you're weighing that tradeoff, [proprietary versus open-source model selection](https://amirteymoori.com/choosing-the-right-llm-proprietary-vs-open-source-in-2025/) is the calculus that decides it.

## How It Stacks Up Against Claude and GPT

Here's the honest version. On SWE-bench Verified, third-party comparisons put V4-Pro around 91%, a few points behind the top Claude Opus results. On the harder SWE-bench Pro, the same comparisons land V4-Pro near 55%, with Claude Opus 4.7 around 64%. Claude Fable 5 has been independently measured near 80% on SWE-bench Pro.

| Model | Weights | Total / active params | API price in / out per Mtok |
|---|---|---|---|
| DeepSeek V4-Pro | MIT, on Hugging Face | 1.6T / 49B | $0.435 / $0.87 |
| DeepSeek V4-Flash-0731 | MIT, on Hugging Face | 284B / 13B | $0.14 / $0.28 |
| Claude Fable 5 | Closed | Not disclosed | $10 / $50 |
| GPT-5.6 family | Closed | Not disclosed | Varies by tier |

Read those numbers with suspicion. Different harnesses, different scaffolds, different retry budgets, and vendors rarely publish the agent loop they used. What I'd say confidently is that V4-Pro clears the bar for routine multi-file refactors and fails more often than Fable 5 on genuinely tangled work, which tracks with the wider [developer model comparison](https://amirteymoori.com/the-5-best-large-language-models-for-developers-in-2025-a-practical-comparison/) I keep updated.

The gap is real. It's also roughly a year wide instead of the three years it used to be.

## The Price Angle

Compare V4-Pro at $0.435 in and $0.87 out against Fable 5 at $10 and $50. Input is about 23x cheaper. Output is about 57x cheaper. Cache hits go further still, quoted at $0.003625 per million for Pro and $0.0028 for Flash, which is roughly a 50x discount over a cache miss.

Numbers like that change what you're willing to build. Agentic loops that re-read a codebase every turn become affordable. Nightly eval pipelines become cheap. I broke the cross-vendor math down in the [2026 LLM API pricing comparison](https://amirteymoori.com/llm-api-pricing-2026-claude-gpt-deepseek-qwen-comparison/), but the short version is that DeepSeek still prices roughly an order of magnitude below the US labs.

Caveat: cheap tokens invite waste. Prompt caching and routing still pay for themselves, and the techniques in my [inference cost guide](https://amirteymoori.com/llm-inference-optimization-how-to-cut-your-ai-costs-by-80-without-sacrificing-quality/) apply here too.

## What "Open Weights" Does and Doesn't Mean

The license is real. Both models are published under MIT on Hugging Face at [deepseek-ai/DeepSeek-V4-Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), with commercial use, modification, redistribution, and fine-tuning permitted without usage caps or revenue thresholds. That's a meaningfully different deal from Llama's community license.

What you don't get is training data, training code, or the RL recipe. Open weights, not open source. You can fine-tune the released checkpoint, the same way I'd [fine-tune a small open model for a narrow task](https://amirteymoori.com/fine-tune-gemma-4-e4b-english-swedish-translation/), but you can't reproduce it.

Then there's hardware. V4-Pro ships in FP4 and FP8 mixed precision and still wants roughly a terabyte of VRAM. Practical single-node deployments start around 8x H200, with Blackwell nodes or a two-node H200 cluster reported for full 1M context. Quantizing down to Q4 leaves you near 430GB of weights before KV cache. There's no quant level that makes Pro a workstation model.

Flash is the one you can actually host. 13B active, 284B total, GGUF builds circulating, and day-zero support in vLLM and SGLang. Still a serious server, but a plausible one.

## The Catch List

Serving complexity is the first tax. Expert-parallel sharding across GPUs with high-bandwidth interconnect is a distributed systems problem, not a `pip install vllm` problem, and your throughput depends on routing balance you don't control.

Privacy is the second. The hosted API is operated from China, and DeepSeek's consumer-facing terms have drawn regulatory attention in several European markets. Handling customer data under GDPR? Read the data processing terms closely, or route through a Western host serving the open weights.

Geopolitics is the third, and it cuts both ways. Export controls and procurement rules leave some organizations unable to touch the hosted API at all. The MIT weights are the escape hatch, which is precisely why they were released that way.

Last one: benchmarks age in weeks. Everything above reflects early August 2026.

## Frequently Asked Questions

### Can I actually download DeepSeek V4?

Yes. Both are published on Hugging Face and ModelScope under the MIT license, in FP8 and FP4-mixed precision. No gated access request, no acceptable-use addendum, no revenue threshold. Quantized community builds for llama.cpp, Ollama, and LM Studio exist for Flash.

### Is V4-Pro better than Claude for coding?

Not in my experience, and the public benchmarks agree. Claude's top models still lead the harder agentic coding evals by a meaningful margin. V4-Pro is close enough that price wins for high-volume, moderately hard work, and loses when one bad multi-file edit costs you an afternoon.

### What hardware do I need to self-host it?

For V4-Pro, plan on roughly 1TB of VRAM, meaning 8x H200 in one node or a multi-node cluster with InfiniBand. For V4-Flash, requirements drop sharply thanks to 13B active parameters, and a well-specced single GPU server handles it. If you don't run a datacenter, use the hosted API or a third-party provider.

### Should I worry about sending data to DeepSeek's API?

Treat it like any offshore processor. For public code, prototypes, and non-sensitive workloads, it's fine. For regulated or customer data, self-host the weights or use a Western provider hosting the same MIT-licensed model, which is a legitimate option precisely because the license permits it.

### How far behind proprietary models is open weight now?

Historically the lag ran 6 to 18 months, and trackers like llm-stats.com show that window shrinking release over release. V4 puts the gap at the short end of that range for coding and agents. On the hardest reasoning tasks, the frontier labs still hold a clear lead.