DeepSeek dropped V4, and the numbers are staggering. A fully open-weight model trained entirely on Huawei Ascend chips, released under a permissive license, delivering GPT-5-class performance at less than one tenth the inference cost. For developers building on LLM APIs, this changes the economics overnight.
The timing matters. US export controls were designed to prevent exactly this, forcing China into a corner on AI hardware. Instead, DeepSeek responded by proving that the software stack and architecture innovations matter more than access to the latest NVIDIA silicon. V4 is the first frontier model that genuinely doesn't need CUDA.
The Numbers That Matter
Open the HuggingFace collection page for deepseek-ai and you'll find four V4 variants:
| Variant | Parameters | Output Price | Downloads |
|---|---|---|---|
| V4 Flash | 158B | $0.20/M tok | 2.24M |
| V4 Flash Base | 292B | Self-host | 97K |
| V4 Pro | 861B | $2.60/M tok | 2.05M |
| V4 Pro Base | 1.6T | Self-host | 24K |
API pricing comparison: DeepSeek V4 Flash at $0.20/M tokens vs GPT-5 at $60/M tokens. Flash is 300x cheaper. Even the flagship V4 Pro at $2.60/M is 23x cheaper than GPT-5.
V4 Flash is shipping 90.4 tokens per second on Fireworks. V4 Pro hits 75 tok/s on Together. These are production throughput numbers, not research paper claims.
For context: if you're burning $1,000/month on GPT-5 API calls, switching to V4 Flash drops that to about $3.30. V4 Pro brings it to $43. That's not a marginal optimization. That's a rewrite-your-cost-model kind of shift.
What Makes V4 Technically Interesting
Every Frontier model in 2026 uses a Mixture of Experts (MoE) architecture. The difference is in the details.
Multi-Head Latent Attention (MLA). DeepSeek introduced MLA with V2 and it's now standard across their lineup. The idea: compress the KV cache into a low-rank latent space during inference, drastically reducing memory usage. For context windows exceeding 128K tokens (which V4 supports), this is what makes serving costs sustainable. Without MLA, the KV cache for a 128K context with 861B parameters would be commercially unviable.
Sparse MoE routing. Only a fraction of experts activate per token. V4 Flash activates roughly 16-20 out of 158, V4 Pro activates about 40-60 out of 861. This is why total parameter count matters less than you'd think. The effective compute per token is much smaller, and that's where the speed and cost advantage comes from.
Huawei CANN stack. This is the geopolitical story. DeepSeek trained V4 on Huawei Ascend 910C accelerators using CANN (Compute Architecture for Neural Networks) instead of CUDA. For years, the narrative was that CUDA's moat was unassailable. DeepSeek just proved otherwise at frontier scale.
Training and deployment architecture: Huawei Ascend hardware layer, MoE + MLA model design, and the multi-provider deployment ecosystem. All open weights, Apache 2.0 licensed.
Getting V4 Running Locally
You don't need a datacenter. Here's how to spin up V4 Flash on consumer hardware.
Option 1: Cloud API (5 minutes)
import openai
client = openai.OpenAI(
base_url="https://api.deepinfra.com/v1/openai",
api_key="your-deepinfra-key"
)
response = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4-Flash",
messages=[{"role": "user", "content": "Explain DPU offloading in 3 sentences."}]
)
print(response.choices[0].message.content)
That's $0.20 per million output tokens. The client code is identical to your existing OpenAI setup. Same schemas, same tool calling interface, same structured output support.
Option 2: Self-Hosted with vLLM
pip install vllm
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash
vllm serve deepseek-ai/DeepSeek-V4-Flash --tensor-parallel-size 4 --max-model-len 131072 --port 8000
V4 Flash fits on 4x A100-80GB or 8x RTX 4090 with quantization. V4 Pro needs more serious hardware (8x H100 minimum for full precision), but GGUF quantized versions are already on HuggingFace for lower-resource setups.
Option 3: Ollama (Simplest)
ollama pull deepseek-v4:flash
ollama run deepseek-v4:flash
Ollama handles quantization and memory management automatically. Not the fastest option, but it works on a single GPU MacBook.
Where V4 Wins (and Where It Doesn't)
From practical testing, here's what I've found:
Clear wins:
- Cost-sensitive production workloads. If you're serving thousands of requests, the 300x price difference from GPT-5 is real money.
- Open-source toolchains. You own the model. No vendor lock-in, no API deprecations, no surprise price hikes.
- Fine-tuning. Full weights means you can actually fine-tune V4 for your domain, unlike GPT-5 where you're limited to the API's fine-tuning surface.
Still maturing:
- Structured output reliability on some providers. DeepInfra has it sorted, but check your specific host.
- Multimodal support. V4 is text-only (unlike GPT-5's vision capabilities). DeepSeek has separate VL and OCR models for vision tasks.
- Ecosystem tooling. LangChain and LlamaIndex work fine, but some edge cases in agent frameworks are still being ironed out.
The Bigger Picture
DeepSeek V4 isn't just another model release. It's the moment where the AI supply chain visibly bifurcated. One track runs on NVIDIA hardware with CUDA and proprietary APIs. The other runs on alternative silicon with open weights and commodity pricing.
For developers, this is unambiguously good. Competition at the frontier drives prices down and keeps weights open. Six months ago, running a GPT-5-class model locally was a fantasy. Today it's a pip install vllm command.
The Huawei story is the wildcard. If Ascend continues improving and DeepSeek keeps executing at this pace, the hardware monopoly that's defined the last five years of AI becomes a lot less relevant. For anyone building on LLMs, that's worth paying attention to.
What's your experience with DeepSeek V4? Are you running it in production yet, or sticking with the incumbents?
Top comments (0)