How far has DeepSeek V4, an open-weight AI that rivals paid models at ~1/30 the cost, come?

DeepSeek V4 is an open-weight AI whose "weights" (the billions of numbers that form its brain) are released for free under the MIT license, and on real coding tests it comes within striking distance of top paid models while costing roughly 1/30 as much per token to use. Because the weights are public, you can download and run it yourself on a capable GPU server or cloud. The catch is that the weights are free, but the server and GPU to run them are not.

In one line: MIT-licensed open weights, top open-weight coding score (SWE-bench 80.6%), and about 1/30 the output cost. It is a real option when skill, cost, and control all matter.

How is "open-weight" different from open-source?#

Start with a term that is easy to confuse. Open-weight means the model's weight files (the billions of numbers learned during training, that is, the model's "brain") are released so you can download and run them yourself. That is different from fully open-source, which would also release the training data and the complete training code. As an analogy, you get the finished lock (the weights) in full, but not necessarily the blueprint and the bill of materials (the data and training pipeline) that produced it.

DeepSeek V4 was released by China's DeepSeek in April 2026, and its MIT license broadly permits commercial use, modification, and redistribution. Three practical freedoms follow: (1) you can self-host it on your own servers, (2) you can fine-tune it on your own data, and (3) you are not locked into one company's API. If an API is discontinued, raises prices, or changes policy, having the weights in hand means you can keep running.

How can it be huge yet cheap and fast?#

The key is a design called Mixture-of-Experts (MoE): the model is enormous, but only a fraction of it actually does the math for each token. Once you see this, "1.6 trillion parameters but cheap" stops sounding like a contradiction.

A normal model engages all of its parameters to produce each answer. MoE is different. Even if a big company has many departmental experts, a single question only calls in a few relevant departments; the rest rest for this token. So the total parameters (overall size) and the active parameters (what is actually computed per token) diverge sharply.

DeepSeek V4 comes in two sizes. V4-Pro (huge) has 1.6 trillion total parameters but activates only about 49 billion per token, while V4-Flash (usually more realistic) has 284 billion total and activates about 13 billion. Compute cost is roughly proportional to the active parameters, which is why they are cheap and fast for their size. On top of that, a 1-million-token context lets them ingest and process a book-length document at once.

How close is it to top paid models?#

Across three axes, skill, cost, and control, it has become a real candidate. Take them one at a time.

Skill. On the standard test of fixing real GitHub bugs, SWE-bench Verified, it scores 80.6%, the top open-weight result and tied with Google's Gemini 3.1 Pro. It still trails top paid models like Claude Opus and GPT-5.5 (around 88%) by a few points, but the gap has narrowed enough to call it striking distance.

Cost. About 1/30 per token on output. Roughly $0.87 per million output tokens versus about $25 for Opus (input and cache prices differ, so compute it against your own usage pattern). The heavier your workload, the wider this gap opens.

Control. Running it on your own infrastructure means you hold the data, cost, and dependencies. And because its API speaks both OpenAI and Anthropic formats, you can point existing tools like Claude Code at a new endpoint and use it right away.

What should you watch out for when self-hosting?#

First, hardware. To run it yourself, Flash is more realistic than Pro, but neither fits a normal laptop; you need a high-end GPU server. Even though the active parameters are small, you must keep all of the total parameters in memory, which demands multiple GPUs and large VRAM.

Second, a quality trap. Many hosts quantize output compute to fp8 (8-bit) to cut cost, which can move results subtly away from the original 16-bit weights. If quality matters, verify the model is served at bf16 (16-bit). The same "DeepSeek V4" can feel different depending on serving precision.

Third, the rest. Chinese models can carry censorship or policy-shaped answers, and you should review data security and licensing before adopting. For long agent loops or multimodal work, top paid models still lead.

What is the takeaway?#

The idea that "frontier AI belongs only to Big Tech's closed models" is wobbling. As freely released weights reach striking distance on coding and reasoning, open-weight models have become a default candidate wherever cost and control matter. For work that absolutely needs the last few points of peak score, top paid models are still the safe choice, but across a wide range beyond that, DeepSeek V4 is a serious alternative. That is the shift Hax is here to watch.

Note: for open-weight models, benchmark scores, per-token prices, and serving precision change fast as of 2026. The figures above are approximations at release time, so re-check the Hugging Face model card and each host's latest pricing and precision notices before adopting.