DeepSeek V4 Pro Deep Dive: A 1.6T-Parameter "Chinese Answer" That Flips the Table at 1/70 the Price
> "When GPT-5.5 doubled its API price to $30/M output tokens, DeepSeek V4 Pro launched the next day at $3.48/M. This isn't coincidence—it's a declaration of war."
On April 24, 2026, DeepSeek released the V4 Pro preview. The same day, seven domestic Chinese chip vendors—Huawei, Cambricon, Hygon, Moore Threads, and others—announced Day 0 adaptation.
This is not an ordinary model iteration. It is a story about engineering efficiency, geopolitics, and open-source conviction.
---
Key specs: V3 vs V4 Pro
| Metric | DeepSeek V3 | DeepSeek V4 Pro | Change | |--------|-------------|-----------------|--------| | Total parameters | 671B | 1.6T | +138% | | Active params/token | 37B | 49B | +32% | | Context window | 128K | 1M | 8x | | NIAH @ 1M | ~45% | 97% | +52pp | | KV Cache | baseline | 7% of V3 | -93% | | License | MIT | MIT | unchanged | | API input price | ~$0.14/M | $1.74/M | adjusted | | API output price | ~$0.28/M | $3.48/M | adjusted |
Core insight: total parameters jumped from 671B to 1.6T, but active parameters only grew from 37B to 49B. This isn't "bigger"—it's "smarter": sparsity converts parameter growth into knowledge capacity, not compute cost.
---
Three engineering breakthroughs
1. Engram: conditional memory system
Traditional Transformers lose recall as context grows (a nominal 1M window may only retrieve ~45%). Engram splits static knowledge retrieval from dynamic reasoning:
- Static knowledge (API signatures, historical facts): O(1) hash lookup against an embedding table
- Dynamic reasoning (logic, code generation): routed through the MoE backbone (256 experts, 8 experts/token)
- Context-Aware Gating decides the path per token, suppressing noise on conflict
- Coding: LiveCodeBench 93.5% is the highest score of any tested model; Codeforces 3206 beats GPT-5.4 (3168)
- Software engineering: near-parity on SWE-Verified, but clearly behind Opus 4.7 on SWE-Pro
- Long context: 97% vs GPT-5.5's 74%—a crushing advantage
- Agentic/terminal tasks: GPT-5.5 leads (Terminal-Bench 82.7%)
- Chinese: SuperCLUE 70.98, ranked #1 domestically
- Training: most likely primarily on NVIDIA H800 (per MIT Technology Review and other third-party analyses); DeepSeek officially cites Ascend 910C + H800 and denies using smuggled Blackwell chips
- Inference: fully adapted to Huawei Ascend 950PR—the only domestically commercial chip explicitly supporting FP4 inference (2 PFLOPS FP4, 112GB HiBL memory, 1.4 TB/s bandwidth)
- Day 0 adaptation by seven vendors: Huawei, Cambricon (open-sourced on GitHub), Hygon, Moore Threads (ModelScope), MetaX, Baidu Kunlun, and others—a first for domestic AI chips
- FP4/FP8 mixed precision: memory down to 1/2–1/4 of FP16, higher throughput
- Three reasoning modes: Non-think / Think High / Think Max (deep thinking still cheaper than competitors' standard modes)
- Context caching discounts repeated input tokens
- Weights: Pro = 865GB (FP4+FP8); Flash = 284B total / 13B active; MIT license, fully commercial
- API: compatible with both OpenAI ChatCompletions and Anthropic formats
- Open-source models can now compete with closed ones at trillion scale
- Sparse architecture + conditional memory appears to be the right path for parameter scaling
- AI pricing power is shifting toward teams with the best engineering efficiency
- China's AI compute ecosystem achieved same-day model adaptation for the first time
- DeepSeek official: https://deepseek.ai, https://deepseek.ai/deepseek-v4
- TrendForce Day 0 adaptation coverage: https://www.trendforce.com/news/2026/04/29/news-huawei-ascend-cambricon-and-hygon-completed-day-0-adaptation-to-deepseek-v4/
- TrendForce Ascend 950PR analysis: https://www.trendforce.com/news/2026/04/07/news-decoding-deepseek-v4-how-huaweis-ascend-950-pr-is-powering-chinas-push-to-break-cuda-dependence/
- The China Academy: https://thechinaacademy.org/why-deepseek-v4-hasnt-fully-cut-ties-with-nvidia/
- Comparison pieces: https://www.datacamp.com/blog/deepseek-v4-vs-gpt-5-5, https://spectrumailab.com/blog/claude-opus-4-7-vs-gpt-5-5-vs-gemini-3-1-pro-vs-deepseek-v4-comparison-2026, https://www.llmreference.com/compare/gpt-5.5/deepseek-v4-pro, and others as listed in the source post
DeepSeek found the optimal split is 20–25% memory / 75–80% compute—pure MoE is suboptimal because GPU cycles get wasted re-retrieving static facts.
Results: MMLU +3.4, BBH +5.0, HumanEval +3.0, and NIAH @ 1M jumping from 84.2% to 97%.
2. mHC: Manifold-Constrained Hyper-Connections
Standard hyper-connections amplify signals ~3000x in deep networks—a primary source of training instability. mHC uses Birkhoff Polytope projection to cap amplification near 2x, at a cost of 6.7% training overhead.
| Configuration | BBH | |---------------|-----| | Baseline | 43.8 | | Unconstrained HC | 48.9 | | mHC | 51.0 |
Without mHC, training at V4 scale would either diverge or require multiples more compute to stabilize.
3. DSA: DeepSeek Sparse Attention
Full attention over 1M tokens is economically infeasible (O(n²)). DSA combines:
1. Lightning Indexer — sublinear context scanning to locate relevant segments 2. Fine-grained token selection into the attention window 3. Hybrid attention: CSA (Chunked Sparse Attention) + HCA (Hierarchical Compressed Attention)
Result: ~50% reduction in long-context compute and KV cache compressed to 7% of V3—making 1M-token inference economically viable.
---
Benchmarks (as of 2026-04-28, per vendor releases and third-party tests)
| Benchmark | V4 Pro | GPT-5.5 | Claude Opus 4.7 | Gemini 3.1 Pro | |-----------|--------|---------|-----------------|----------------| | MMLU Pro | 87.5 | 88.1 | – | – | | SWE-bench Verified | 80.6 | 88.7 | 80.9 | – | | SWE-bench Pro | 55.4 | – | 64.3 | – | | LiveCodeBench | 93.5 | – | – | – | | Codeforces | 3206 | – | – | – | | NIAH @ 1M | 97% | 74% (512K–1M) | – | – | | Chatbot Arena | 1460 | 1488 | – | – |
Takeaways:
---
Pricing: flipping the table
| Model | Input (\(/M) | Output (\)/M) | Context | |-------|------------|--------------|---------| | DeepSeek V4-Flash | $0.14 | $0.28 | 1M | | DeepSeek V4-Pro | $1.74 | $3.48 | 1M | | GPT-5.5 | $5.00 | $30.00 | ~1M | | Claude Opus 4.7 | $5.00 | $25.00 | 1M | | Gemini 3.1 Pro | $2.00 | $12.00 | 200K |
V4-Pro's output price is ~1/8.6 of GPT-5.5's and ~1/7.2 of Opus 4.7's—an 88% saving at any volume. Note the launch price is a 75% discount; list price is expected around $2.90/$5.80, still ~5x cheaper than GPT-5.5 output. DeepSeek's pricing page also states that once Huawei Ascend 950 supernodes ship at scale in H2, Pro prices will drop significantly.
---
Hardware ecosystem: NVIDIA vs Ascend
Despite prioritizing Huawei, V4 weights still run on NVIDIA GPUs via vLLM/SGLang. Cutting NVIDIA would effectively close-source the model for the 99% of developers without Huawei hardware—optionality is worth more than any single hardware relationship.
---
Other design details
---
Known weaknesses
1. Multimodality: clearly behind GPT-5.5 and Gemini 3.1 Pro 2. Hallucination rate: preview instability; Think modes needed to suppress it (SimpleQA-Verified 57.9 vs Gemini's 75.6) 3. SWE-Pro gap: 55.4% vs Opus 4.7's 64.3% 4. API stability: occasional 500 errors in the preview
---
Strategic significance
When GPT-5.5 doubled prices, the signal was: "frontier AI is a luxury." DeepSeek V4 Pro answered at $3.48/M: "frontier AI can be infrastructure."
---