In the quiet arithmetic of the AI arms race, DeepSeek has always argued that efficiency is a more durable weapon than raw scale. With the release of DeepSeek-V4, China's most consequential AI lab has proved its theorem — and in the process laid a charge beneath the foundations of Silicon Valley's financial narrative.
Released in a daytime drop that broke the lab's legendary "midnight launch" tradition, DeepSeek-V4 arrives not merely as a new model but as a formal declaration: the open-source world can match — and in critical dimensions surpass — the closed-source dominance of OpenAI, Anthropic, and Google, and it can do so on hardware the United States set out to deny China entirely.
The Architecture of Defiance
DeepSeek-V4 ships in two configurations that together define a new tier of capability. The flagship DeepSeek-V4-Pro carries a staggering 1.6 trillion total parameters, 49 billion of which are activated during inference. Its leaner sibling, DeepSeek-V4-Flash, holds 284 billion total with just 13 billion activated. Both support a native context window of one million tokens — a claim, unlike many in AI, backed by genuine engineering rather than marketing arithmetic.
The headline achievement is the hybrid attention architecture, a marriage of Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Where most transformer models choke on the memory demands of million-token contexts, DeepSeek's hybrid compresses vast blocks of the KV cache into compact summary representations — reading a document the way a skilled intelligence analyst reads a briefing, not the way a court stenographer transcribes every word. The result is extraordinary: at one million tokens, V4-Pro requires only 27% of the inference FLOPs and 10% of the KV cache that DeepSeek-V3.2 consumed. V4-Flash achieves even more dramatic compression — 10% of FLOPs and 7% of KV cache against its predecessor.
The architecture extends inward as well. Manifold-Constrained Hyper-Connections (mHC) replace conventional residual pathways with streams whose mixing is constrained to the doubly stochastic matrix manifold, enforced by Sinkhorn-Knopp projection — a technique borrowed from convex optimization to prevent unstable signal amplification at this depth. The Muon optimizer, replacing AdamW for most parameters, rounds out the trifecta of architectural novelty.
Post-training was equally radical. Rather than lean on the industry-standard mixed reinforcement learning, DeepSeek cultivated independent domain experts — specialists in coding, mathematics, and reasoning — and then merged their knowledge through On-Policy Distillation: more than ten teacher models distilled into a single student along the student's own trajectories. The method sidesteps the mode collapse and capability dilution that have plagued weight-merging approaches, and its results are visible across the benchmarks.
The Scoreboard
In maximum reasoning mode — V4-Pro-Max — DeepSeek claims a Codeforces rating of 3206, placing it alongside, or fractionally above, GPT-5.4 (3168) and Gemini-3.1-Pro (3052) by the lab's own evaluation. On LiveCodeBench, V4-Pro-Max reaches 93.5%, the highest reported score in the comparison set. On IMOAnswerBench, it scores 89.8%, trailing only GPT-5.4's 91.4%. Most strikingly, in a hybrid informal–formal verification pipeline, the model achieved 120 out of 120 on the Putnam-2025 mathematics competition.
Long-context performance tells a more nuanced story. On MRCR at one million tokens, V4-Pro-Max scores 83.5%, outperforming Gemini-3.1-Pro (76.3%) but trailing Claude Opus 4.6 (92.9%). The paper itself discloses that performance degrades beyond 128K tokens — a crucial admission. One-million-token support does not mean one-million-token perfection. It means one-million-token viability, which is an entirely different and still remarkable claim.
The agentic benchmarks show where the frontier remains closed. On Terminal Bench 2.0, GPT-5.4 leads at 75.1%; Gemini-3.1-Pro reaches 68.5%; V4-Pro-Max lands at 67.9%. The gap is narrowing, but it has not closed.
The Hardware Revelation
For Western observers, the most politically charged dimension of V4's release is not its benchmark scores — it is the silicon on which it was trained. Developers have confirmed that V4 was produced on Huawei's Ascend 950 chips and the CANN software ecosystem, hardware that US export controls were explicitly designed to keep China from accessing at competitive capability levels.
The implication is direct and uncomfortable: the chip bans did not stall Chinese AI development. They catalyzed it. DeepSeek has now achieved Opus 4.6-class open-weight performance on restricted domestic silicon, and as developers around the world flock to the model, Huawei's ecosystem matures under production load. The financial premise of maintaining a technology gap through export restriction has, at minimum, been badly complicated.
CTOL Digital Solutions' House Review
DeepSeek-V4 is best read not as a benchmark release but as a full-stack systems engineering paper masquerading as a model card. The real contribution is a plausible, reproducible recipe for making million-token intelligence routinely serviceable in production — combining compressed/hybrid attention, MoE routing efficiency, FP4-aware quantization, deterministic custom kernels, and on-disk KV-cache prefix reuse into a coherent deployment primitive.
The most underappreciated contributions are infrastructural: batch-invariant deterministic kernels for reproducible post-training; DSec sandboxing that manages hundreds of thousands of concurrent agentic environments; interleaved thinking that preserves reasoning chains across tool calls; and on-disk KV reuse that turns shared prefixes from a memory liability into a cache asset. This is not a paper about a smarter model. It is a paper about a cheaper production system.
Four academic trajectories will be reshaped by it: long-context attention economics (CSA/HCA as a reference architecture), residual-path design as a serious scaling axis (mHC), multi-teacher post-training consolidation (OPD replacing mixed RL), and hybrid informal–formal mathematical reasoning pipelines.
Developer response has divided sharply between admiration and qualified critique. The community has widely praised DeepSeek's technical report for its unprecedented transparency — openly acknowledging a 3–6 month world-knowledge lag behind Gemini, openly disclosing long-context degradation beyond 128K, and openly rating the model below Opus 4.6 Thinking in the lab's own internal engineering survey of 85 developers, where only 52% named V4-Pro their default coding model.
On code quality, the verdict is "powerful but unrefined." V4-Pro targets root causes with the diagnostic precision of Opus and GPT-5.4 — correctly identifying obscure bugs rather than guessing at them — but the generated architecture and UI outputs lack the aesthetic elegance that has become Claude Opus's signature. In "Think High" mode on complex multi-file projects, the model intermittently loses coherence and skips implementation steps; developers report needing to escalate to "Think Max" to achieve reliable zero-shot completion.
On economics: V4-Flash is the commercial sleeper. At roughly $0.14 per million tokens — about 1 RMB — it is positioned squarely for the high-volume bulk workflows that define real enterprise AI spend. V4-Pro-Max, at roughly $3.30 per million tokens, sits at the high end of open-weight pricing. The speed tradeoff is real; heavy reasoning modes produce measurably slower outputs.
CTOL Digital Solutions' House Conclusions
This release is more than a model milestone. It is a statement of strategic sovereignty. DeepSeek-V4 advances China's objective of building a complete, end-to-end AI stack — domestic hardware, domestic software, domestic frontier LLM — with training and inference self-sufficiency that is no longer aspirational but demonstrably functional.
Wenfeng Liang's lab continues to operate under a different economic logic than its Western counterparts. Where US frontier labs face investor pressure to close their weights and monetize access, DeepSeek has again chosen open release under the MIT License — a decision that accelerates global AI capability diffusion and deliberately expands the frontier the entire research community can build upon.
The gap to Claude Opus 4.6 is real. The lab says so directly. But the trajectory is unambiguous: with each release, the open-source frontier absorbs what was previously a closed-source moat. DeepSeek-V4 does not beat every closed model. It narrows the gap on every dimension that matters commercially, while attacking the single most expensive bottleneck in frontier AI deployment — the cost of processing long context at scale.
We congratulate DeepSeek on this release, and recognize Wenfeng Liang's continued commitment to advancing the global open-source LLM commons rather than yielding to the gravitational pull of walled-garden monetization.
Sources: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro