Key Takeaways
- DeepSeek V4 has ~1.6 trillion total parameters (49B active per token via MoE), a 1 million token context window, and runs on Huawei Ascend 950PR chips. Not NVIDIA. Not AMD. DeepSeek deliberately gave Chinese chipmakers early optimization access while denying NVIDIA and AMD.
- API pricing is up to 10-50x cheaper than GPT-5.4 on the V4-Flash tier (V4-Pro is roughly 1.5x cheaper). For agentic AI systems that make dozens or hundreds of LLM calls per task, this changes which workflows are economically viable.
- Mixture of Experts (MoE) architecture activates only a small fraction of parameters per token (V4-Pro: ~49B of 1.6T). This is how V4 achieves trillion-parameter capability without trillion-parameter compute costs.
- Open-weight models at this capability level make self-hosted AI viable. If you don't want your data leaving your network, self-hosted DeepSeek V4 at $0 in API costs is a real option.
- The model layer is the most replaceable part of any AI coding stack. The tooling, context management, and prompt engineering around the model often matter more than which LLM is underneath.
DeepSeek V4 shipped in April 2026. It has roughly 1.6 trillion total parameters, a 1 million token context window, native multimodal support across text, images, video, and audio, and it runs on Huawei's Ascend 950PR processor. Not NVIDIA. Not AMD. Huawei.
Reuters confirmed the chip detail on April 3. According to The Information, DeepSeek deliberately gave Chinese chipmakers early optimization access while denying NVIDIA and AMD that window. That decision turns V4 from a model launch into a geopolitical statement about the future of AI compute infrastructure.
The full V4-Pro released in April 2026, with V4-Flash (a smaller 284B-total/13B-active variant) also offered on DeepSeek's API. A smaller variant running ahead of the full release is a pattern DeepSeek has used before, and it signals the full model is close.
What 1.6 Trillion Parameters Actually Means
V4 uses a Mixture-of-Experts (MoE) architecture. V4-Pro contains roughly 1.6 trillion total parameters, but only about 49 billion activate for any given token (V4-Flash is 284B total / 13B active). This is the same approach DeepSeek used with V3.2, which had 671 billion total parameters.
The practical effect: inference costs stay roughly flat even as the model gets dramatically larger. You get the capability of a 1.6-trillion-parameter model while only paying for the compute of a 49-billion-parameter one. That is why DeepSeek's pricing has been consistently aggressive. They are not subsidizing cheap prices. The architecture is genuinely cheaper to run.
For comparison, nobody outside OpenAI and Anthropic knows exactly how many parameters GPT-5.4 or Claude Opus 4.5 use. But their API pricing tells the story:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 |
| DeepSeek V4-Pro | $1.74 | $3.48 |
| GPT-5.4 | $2.50 | $15.00 |
| Claude Opus 4.5 | $5.00 | $25.00 |
That is not a marginal difference. V4-Flash runs roughly 10-50x cheaper than GPT-5.4 on comparable tokens (about 18x on input, ~54x on output). V4-Pro is roughly 1.5x cheaper — and DeepSeek cut V4-Pro prices ~75% further on May 31, 2026, widening the gap.
Three Architectural Breakthroughs
The raw parameter count matters less than three specific engineering decisions that make V4 possible at this scale.
Engram Conditional Memory
The hardest problem with million-token context windows is retrieval accuracy. Stuffing a million tokens into the context is useless if the model can't find the right needle in that haystack.
DeepSeek published research on Engram, a conditional memory system from January 2026 that separates static facts (API signatures, database schemas, configuration patterns) from dynamic reasoning. On the Needle-in-a-Haystack benchmark at 1 million tokens, the paper reports accuracy improving from 84.2% to 97%. However, Engram was not carried into the shipped V4 architecture — the released model uses a different long-context approach, so the paper should be read as research direction rather than a description of V4 itself.
That is the difference between a model that technically supports long context and one that actually uses it reliably.
Training Stability at Trillion-Parameter Scale
Training instability is a real problem at trillion-parameter scale. Gradients explode. Loss curves spike. Entire training runs fail. DeepSeek's V4 technical report details how the model was trained at this scale on Ascend hardware. The headline trade-off is the same one every MoE lab faces: you only realize the efficiency win if you can actually stabilize training at this size.
Sparse Attention for Long Context
Standard dense attention across 1 million tokens would make inference prohibitively slow. V4 uses sparse attention mechanisms so that the model can serve a 1 million token window at practical speeds on inference hardware that is not top-of-the-line NVIDIA. The specific mechanics are described in DeepSeek's technical documentation rather than in independent third-party analyses.
The Huawei Chip Story
This is the part that matters beyond benchmarks.
US export controls have cut Chinese AI labs off from NVIDIA's most advanced GPUs. The H200, the B300, none of these ship to China legally. The chips China can legally buy, like the H20, are significantly weaker.
Huawei's Ascend 950PR is the answer. According to TrendForce, the chip delivers roughly 2.87x the compute performance of the NVIDIA H20 and integrates 112GB of Huawei's in-house HiBL memory at 1.4 TB/s bandwidth. It is manufactured by SMIC using their N+3 process, which puts it roughly at 5nm-class performance.
The chip launched in Q1 2026, and the demand surge has been immediate. Reuters reports that Alibaba, ByteDance, and Tencent have placed orders totaling hundreds of thousands of units. EE Times China reports chip prices jumped roughly 20% on the back of V4 anticipation alone.
Huawei planned to ship around 750,000 Ascend 950PR chips in 2026 (per Reuters), with mass production ramping from April and full shipments expected in the second half of the year. The 950PR is already shipping. Earlier Ascend 910C production plans and next-generation Ascend 960/970 chips are less firmly documented and should be treated as less certain.
If DeepSeek succeeds in running both training and inference on Ascend chips within the next year or two, and stabilizes the full software stack (compilers, operators, communication libraries, distributed training, inference frameworks), the entire model development pipeline becomes independent of CUDA.
That is not a hypothetical scenario. It is a stated objective being actively engineered.
Benchmarks, With Caveats
Leaked internal benchmarks put V4 in competitive range with Western frontier models:
| Benchmark | DeepSeek V4 (claimed) | GPT-5.x family | Claude Opus 4.5 | DeepSeek V3.2 |
|---|---|---|---|---|
| SWE-bench Verified | >80% (vendor claim) | ~72.8% (GPT-5.2 Codex, independent) | ~76.8% (independent harness) | ~70% |
| Context Window | 1M tokens | 1.05M | 200K | 128K |
These numbers are unverified by independent third parties. DeepSeek's internal benchmarks have historically been directionally accurate but somewhat optimistic. Note that on the independent SWE-bench Verified leaderboard, models cluster much tighter (roughly 70-77%) than vendor-reported figures like Opus 4.5's 80.9%. Wait for independent testing before making infrastructure decisions based on these.
The context window is the clearest advantage on paper. At 1 million tokens, V4 can hold roughly 25,000 to 30,000 lines of code or 15 to 20 full-length novels in a single prompt. GPT-5.4 supports a 1.05M-token context. Claude Opus 4.5 at 200K. If the retrieval accuracy holds up in practice, the practical gap is still meaningful for long-context workloads.
Open Source, Open Weights
DeepSeek released V4's weights under the MIT license. They did this with V3, and the community has come to expect it. Thanks to MoE efficiency and quantization (INT8/INT4), smaller V4 variants should be runnable locally on consumer hardware. Dual RTX 4090s or a single RTX 5090 are the commonly cited minimum specs for running larger open-weight models at home.
We wrote about why open-source AI matters for small teams back in March. DeepSeek V4 landing open-weight at this capability level adds another data point. If you are building internal tools, running code review, or doing document analysis, and you do not want your data leaving your network, self-hosted DeepSeek V4 at $0 in API costs is a real option.
For developers already using AI coding tools, V4's positioning is interesting. Claude Code currently leads independent rankings with 80.8% on SWE-bench Verified (vendor-reported; ~75.6% on the independent mini-SWE-agent harness). GPT-5.4 Codex is close behind. DeepSeek V4 claims parity, and even if the real number is a few points lower, the cost difference makes it compelling for teams running large volumes of API calls. You can pair it with open-source coding agents like OpenCode or Aider and get most of the capability for a fraction of the cost.
We covered how Claude Code's internals work after the source leak in April. The model layer is the most replaceable part of any AI coding stack. The tooling, context management, and prompt engineering around the model often matter more than which LLM is underneath. If V4 delivers on the benchmarks, swapping it into existing agent workflows will be straightforward.
What This Means for AI Development
The bigger picture is that frontier AI development is no longer a single-track race. For the past five years, every major model ran on NVIDIA GPUs using CUDA. Training clusters were built around A100s, then H100s, then H200s and B300s. The assumption was that access to the best NVIDIA silicon was a prerequisite for building competitive models.
V4 challenges that assumption directly. A trillion-parameter MoE model, trained and optimized for non-NVIDIA hardware, performing at frontier level. If the benchmarks hold under independent testing, the implication is clear: export controls are not stopping Chinese AI development. They are accelerating the development of an alternative hardware ecosystem.
For builders outside the US-China competition, this creates options. More hardware suppliers means more competition on price. More open-weight frontier models means more choices for self-hosting. The cost of running advanced AI drops for everyone.
We noted in our guide to agentic AI that the biggest bottleneck for agent deployment is inference cost. Agents make dozens or hundreds of LLM calls per task. At GPT-5.4 pricing, that adds up fast. At V4-Flash pricing, you can run roughly 10 to 50 times more agent calls for the same budget. That changes which workflows are economically viable.
The Timing Question
V4 was delayed twice before its April 2026 launch. The full V4-Pro and the smaller V4-Flash are both now live on DeepSeek's API.
The strongest signal before launch was V4-Flash running on API infrastructure ahead of the full release. When DeepSeek starts stress-testing infrastructure variants, a full launch usually follows within weeks.
If you are making decisions about AI model providers for current projects, the smart move is to look at the benchmarks. Not the vendor-reported ones. The independent ones from the community, from SWE-bench runners, from the developers who have put V4 through real codebases since April.
The specs are ambitious. The pricing is disruptive. The hardware story is geopolitically significant. The question now is how the shipped model holds up under independent testing across real production workloads.
At AWZ Digital, we build AI-powered products for clients across the Middle East and Southeast Asia, and model selection is one of the first architecture decisions we make. Whether it is GPT-5.4 for reasoning-heavy workflows, Claude for code generation, or a cost-optimized stack using open-weight models like DeepSeek, the choice depends on the use case, the data sensitivity, and the budget. If you are evaluating AI models for your next project, talk to us.
