⚡ PortableMind — Offline AI on a USB. Voice, Vision & Chat. No Cloud. No Subscription. Starting at $49 →

DeepSeek V4: Everything We Know About the 1-Trillion-Parameter Model That Could Reshape AI

Analysis 2026-04-22 5 min read By Q4KM

DeepSeek V4 hasn't launched yet. But it might be the most consequential AI model release of 2026 — and the details surfacing suggest it's going to change the competitive landscape in ways that go far beyond benchmark scores.

Here's what's confirmed, what's speculation, and why this launch matters more than most.

What We Know for Sure

Reuters confirmed on April 4, 2026 that DeepSeek V4 will run on Huawei's Ascend 950PR chips — not NVIDIA, not AMD. That makes V4 the first frontier AI model built from the ground up for Chinese semiconductor infrastructure. DeepSeek reportedly gave Chinese chipmakers early access while deliberately denying that window to NVIDIA.

The ripple effects are already visible. Alibaba, ByteDance, and Tencent have placed bulk orders for hundreds of thousands of Ascend 950PR chips, driving prices up 20% in weeks.

Beyond the hardware story, here's what's been confirmed about the model itself:

Spec Detail
Total parameters ~1 trillion (Mixture of Experts)
Active parameters per token ~37 billion
Context window 1 million tokens
Modality Text + image + video (native multimodal)
Training cost ~$5.2 million
License Expected open-source (MIT or Apache 2.0)
Hardware Huawei Ascend 950PR + Cambricon chips
Variants V4 (full), V4-Lite (already in API testing)

The Mixture of Experts design is the key architectural choice. Only ~37 billion parameters activate per response, which means V4 runs more like a 37B model in practice while having access to a trillion parameters of knowledge. It's the same approach behind Llama 4 Maverick (400B total, 17B active), but at nearly three times the scale.

The Timeline: Three Delays and Counting

DeepSeek V4 has been delayed three times, and nobody outside the lab knows the actual target date.

Expected Date What Happened
Mid-February 2026 No launch. No explanation.
March 2026 V4-Lite appeared on API nodes March 9 instead of the full model.
Early-to-mid April 2026 Pre-training confirmed complete. Still no model card or HuggingFace drop.
Late April 2026 Current community expectation. No official confirmation.

The strongest signal that a launch is imminent: V4-Lite has been live on API nodes since early April, with developers reporting a 30% inference speed increase and dramatically improved context recall — 94% at 128K tokens, up from 45% on V3.

One note of caution: a stealth model called "Hunter Alpha" appeared on OpenRouter and was widely attributed to DeepSeek. It was later fingerprinted and confirmed to be Xiaomi's MiMo-V2-Pro, not DeepSeek V4. As of April 22, no verifiable V4 model exists on any public platform.

Our best guess: the last two weeks of April remain plausible, but a slip into early May would fit the pattern.

Why DeepSeek V4 Matters (Beyond Benchmarks)

The Huawei Chip Story

This is the angle most coverage misses. DeepSeek isn't just building a model — they're proving a point about AI sovereignty.

By running V4 on Huawei chips instead of NVIDIA hardware, DeepSeek is demonstrating that frontier AI development doesn't require American semiconductors. If V4 matches or beats GPT-5.4 and Claude Opus 4.7 on key benchmarks while running on Chinese infrastructure, it fundamentally undermines the export control strategy that's been central to US AI policy.

The $5.2 million training cost — if accurate — compounds this. You don't need a billion-dollar compute cluster to build frontier AI. You need clever architecture and efficient training.

The Open Source Angle

DeepSeek has historically released models under MIT or Apache 2.0 licenses. If V4 continues that tradition, it would be the first open-source trillion-parameter model. That's a bigger deal than it sounds:

The Competitive Context

V4 launches into a crowded field. Here's where the April 2026 landscape stands:

Model Parameters Open Source Standout Strength
Claude Opus 4.7 Undisclosed No Coding, agents
GPT-5.4 Undisclosed No Computer use, reasoning
Gemini 3.1 Ultra Undisclosed No Multimodal, long context
Llama 4 Maverick 400B (17B active) Yes Multilingual MoE
DeepSeek V3 671B (37B active) Yes Cost efficiency
DeepSeek V4 (expected) ~1T (37B active) Expected yes Full multimodal + open source

If V4 delivers on its specs, it would sit alone at the intersection of frontier performance, open-source availability, and multimodal capability. No other model checks all three boxes.

What V4-Lite Tells Us

V4-Lite — the smaller variant already in testing — offers a preview of what the full model might deliver:

If these improvements scale proportionally to the full V4, the model could legitimately challenge GPT-5.4 and Claude Opus 4.7 on most benchmarks while being free to download and run.

What to Watch For

When DeepSeek V4 does drop, here's what matters:

  1. The benchmark numbers — Does it actually match GPT-5.4 on GPQA and SWE-bench?
  2. The HuggingFace reception — Download counts and community forks in the first 48 hours
  3. The license terms — MIT/Apache 2.0 would be ideal; any restrictions matter
  4. The Huawei chip performance — Does running on Ascend actually work at scale?
  5. The pricing — If there's an API, how does it compare to OpenAI/Anthropic/Google?

We'll update this post the moment V4 lands. The AI landscape is about to get significantly more interesting.


Last updated: April 22, 2026. DeepSeek V4 remains unreleased as of this writing.

Get these models on a hard drive

Skip the downloads. Browse our catalog of 985+ commercially-licensed AI models, available pre-loaded on high-speed drives.

Browse Model Catalog