Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
TL;DR: DeepSeek V4 launched March 3, 2026, with 1 trillion total parameters (32B active per token via Mixture of Experts), native multimodal support across text, image, and video, a 1 million token context window, and inference pricing of $0.10–$0.30 per million tokens — roughly 50x cheaper than GPT-5.2. Weights are open. It runs on Huawei Ascend chips. It was timed to drop one day before China's politically significant "Two Sessions" legislative gathering. This is DeepSeek's third consecutive moment that Western AI incumbents did not see coming.
DeepSeek V4 is a 1 trillion total parameter large language model released March 3, 2026, by DeepSeek, the Chinese AI lab backed by quantitative hedge fund High-Flyer. It uses a Mixture of Experts (MoE) architecture, releases weights publicly under a permissive license, and ships with native multimodal capabilities, a 1M token context window, and an API pricing structure that undercuts every major Western provider by a factor of 10 to 50.
Those facts taken individually are each significant. Taken together, they describe a model that is attempting to do three things at once: match frontier capability, blow up frontier pricing, and do it on hardware that circumvents U.S. export controls.
To understand why this matters, it helps to understand what came before it. DeepSeek R1, released in January 2025, demonstrated that a Chinese lab with constrained access to cutting-edge NVIDIA chips could produce a reasoning model competitive with OpenAI's o1 — and release it open-source at a fraction of the cost. That was the first "Sputnik moment." DeepSeek V3, released later in 2025, extended the capability benchmark matching into general-purpose tasks, again at aggressively low pricing. V4 is the third iteration of this playbook, but it is a qualitative step up: a model that is no longer trying to match last year's Western frontier but is competing directly with GPT-5.2 and Gemini 2.5 Pro — models released in early 2026.
The open-source dimension is not incidental. It is the strategic core. By releasing weights publicly, DeepSeek prevents competitors from using closed-API pricing and proprietary access as moats. Any organization with sufficient GPU infrastructure can download V4's weights, run inference internally, and avoid per-token API costs entirely. For enterprises processing millions of tokens per day, this changes the economics of AI deployment fundamentally.
The 1 trillion parameter figure that headlines V4's launch requires immediate qualification, because the headline number is not the operationally relevant number.
DeepSeek V4 uses a Mixture of Experts architecture in which the model contains 1 trillion total parameters distributed across many expert sub-networks, but only 32 billion parameters are active on any given forward pass. The MoE routing mechanism selects which expert networks are activated for each input token; the others are dormant during that computation.
This design produces a model that behaves as if it has access to the breadth of a 1 trillion parameter network's learned representations, but runs inference at the computational cost of a much smaller active parameter count. For reference, a dense 32B parameter model — one where all 32B parameters are active for every token — costs roughly the same compute per token as V4's active-parameter footprint.
| Model | Total Parameters | Active Parameters per Token | Architecture |
|---|---|---|---|
| DeepSeek V4 | 1T | 32B | MoE |
| DeepSeek V3 | ~671B | ~37B | MoE |
| GPT-5.2 | Undisclosed | Undisclosed | Dense/Undisclosed |
| Gemini 2.5 Pro | Undisclosed | Undisclosed | MoE (reported) |
| Llama 3.1 405B | 405B | 405B | Dense |
The practical implication is that V4 achieves scale-equivalent capability while running inference at a cost structure closer to a 32B dense model than a 1T dense model. This is why the $0.10–$0.30/M token pricing is achievable — it is not subsidized or predatory in the traditional sense. The MoE architecture genuinely reduces the compute required per token.
DeepSeek's engineering team has refined MoE training and inference over multiple model generations. The routing mechanisms, load balancing across experts, and training stability improvements that went into V3 carried forward into V4 with additional optimization for the Huawei Ascend hardware stack. The efficiency gains are real, measurable, and reproducible — they are not a Chinese government subsidy disguised as cost reduction.
DeepSeek V4 ships with native multimodal capabilities across text, image, and video — not as bolt-on modules or post-training adaptations, but as natively trained modalities. This is a meaningful architectural distinction.
Bolt-on multimodality, as seen in many first-generation multimodal systems, means training a language model on text and then connecting a separately trained vision encoder to the language model's embedding space. The visual understanding is limited to what the vision encoder captures and what the language model's residual stream can incorporate. Native multimodal training means the model processes all modalities from the same fundamental representations, enabling deeper cross-modal reasoning.
For text and image, V4's native multimodality means it can reason about images in ways that draw on the same representations as its text reasoning — not just "describe what you see" but genuine cross-modal inference. For video, V4 can process sequences of frames as a coherent temporal input rather than treating video as a collection of independent images.
The practical applications are immediate: document analysis that reasons across text and embedded charts simultaneously, video understanding for surveillance and content moderation at scale, scientific paper analysis where diagrams and equations are co-processed, and code review workflows where screenshots of UI behavior are analyzed alongside the source code that produced them.
The video capability is the most differentiated. Most frontier models as of early 2026 treat video as a premium add-on with significant latency and pricing overhead. V4 treats it as a first-class input modality in its standard API. At $0.10–$0.30/M tokens — even if video tokens carry a multiplier — this opens video-scale AI analysis to organizations that previously could not afford it.
The 1 million token context window is the specification that will matter most for enterprise deployments, and it is the most underappreciated technical achievement in V4's announcement.
For reference:
| Model | Context Window |
|---|---|
| DeepSeek V4 | 1M tokens |
| Claude 3.7 Sonnet | 200K tokens |
| GPT-5.2 | 128K tokens (standard) |
| Gemini 2.5 Pro | 1M tokens |
| Llama 3.1 405B | 128K tokens |
DeepSeek V4 matches Gemini 2.5 Pro's context ceiling and significantly exceeds GPT-5.2's standard offering. At 1M tokens, the context window is large enough to hold approximately 750,000 words — the equivalent of seven average-length novels, or a complete codebase of moderate size.
What does this unlock in practice? Legal analysis of complete contract archives without chunking. Codebase-wide refactoring that understands the full dependency graph. Scientific literature reviews that process dozens of papers in a single context. Customer support systems that hold the complete history of a long-standing enterprise relationship. Financial analysis that processes full annual reports, 10-Ks, and earnings transcripts for multiple quarters simultaneously.
The technical challenge of a 1M token context window is managing the quadratic scaling of attention computation. Standard transformer attention is O(n²) in context length — doubling the context quadruples the computation. Achieving functional, performant attention at 1M tokens requires one or more of: sparse attention mechanisms, sliding window attention, linear attention approximations, or hardware-level optimizations. DeepSeek's specific approach for V4 has not been fully disclosed as of launch, but the model's MoE architecture and Huawei Ascend optimization suggest it incorporates efficiency mechanisms at both the architectural and hardware levels.
The number that will restructure enterprise AI budget conversations in Q2 2026 is $0.10–$0.30 per million tokens.
For context on what this means against the current pricing landscape:
| Model | Input Price (per M tokens) | Output Price (per M tokens) |
|---|---|---|
| DeepSeek V4 | ~$0.10 | ~$0.30 |
| DeepSeek V3 | $0.27 | $1.10 |
| GPT-5.2 | ~$5.00 | ~$15.00 |
| Claude 3.7 Sonnet | $3.00 | $15.00 |
| Gemini 2.5 Pro | $1.25 | $5.00 |
| Llama 3.1 405B (self-hosted) | Infrastructure cost only | Infrastructure cost only |
DeepSeek V4 is approximately 50x cheaper than GPT-5.2 on output tokens. It is roughly 17x cheaper than Claude 3.7 Sonnet on output. Even against Gemini 2.5 Pro, which has been positioned as a value-competitive option, V4 is approximately 17x cheaper on output.
These are not rounding errors. They are order-of-magnitude differences that change the economics of AI deployment for cost-sensitive workloads. At GPT-5.2's pricing, processing 10 billion tokens per month — a realistic volume for a mid-size enterprise with significant AI automation — costs approximately $150,000/month on output alone. At V4's pricing, the same volume costs approximately $3,000/month.
The implication for AI budget allocation is direct: organizations that have been rationing GPT-5.2 usage to high-value tasks because of cost constraints can now run the same or comparable quality across a far broader set of tasks. The "AI for everything" aspiration that has been economically constrained becomes economically viable.
The pricing also sets a new floor expectation. Enterprise procurement teams negotiating with OpenAI, Anthropic, and Google will now cite DeepSeek V4's pricing as a baseline. This is the same dynamic that played out after V1 and V3 — Western providers respond with pricing pressure or new cost-optimized tiers. V4 accelerates that cycle.
DeepSeek V4's open weights release would be a significant event regardless of the hardware it runs on. The fact that it is specifically optimized for Huawei Ascend chips makes it geopolitically significant in a way that purely technical coverage underplays.
The U.S. government's export control regime — specifically the restrictions on exporting NVIDIA H100 and A100 chips to China — was premised on the assumption that cutting-edge AI capability requires cutting-edge NVIDIA silicon. DeepSeek has now, across three model generations, demonstrated that this assumption is contestable. V4's Ascend optimization is the most direct challenge yet: a model explicitly engineered to run at frontier capability on hardware that the U.S. export regime does not restrict.
Huawei's Ascend 910B chip, the primary target for V4's optimized inference stack, is manufactured in China without NVIDIA technology. Its raw performance per chip is below NVIDIA's H100 in standard benchmarks, but DeepSeek's software optimization — achieved through a combination of custom CUDA-equivalent kernels, MoE architecture that reduces per-token active compute, and infrastructure-level batching improvements — substantially closes the gap for inference workloads.
The open weights release amplifies this geopolitical dimension. Organizations outside China that download V4's weights can run inference on whatever hardware they have access to, including NVIDIA hardware. But organizations inside China, or organizations anywhere with access to Ascend infrastructure, now have a frontier model explicitly optimized for that stack. The export control architecture was designed to prevent frontier model training in China; it did not successfully prevent frontier model development.
The implications for the U.S.-China AI competition narrative are significant. The dominant Western framing — that chip export controls would create a durable capability gap — has been challenged by each DeepSeek release. V4 does not definitively resolve the question of whether the gap has been eliminated, but it substantially raises the evidentiary bar for anyone claiming the controls have been effective.
DeepSeek V4 was released on March 3, 2026. China's "Two Sessions" — the annual plenary sessions of the National People's Congress and the Chinese People's Political Consultative Conference — began March 4, 2026.
This is not a coincidence.
The Two Sessions is China's most significant political event of the year. It is where the State Council sets economic targets, announces major policy priorities, and signals the government's agenda for the year ahead. Technology and AI have been explicit priorities in recent Two Sessions gatherings, with the government announcing support packages, development frameworks, and national AI strategies.
A frontier model release one day before Two Sessions serves multiple functions. It provides concrete evidence of Chinese AI capability at the moment political leadership is assessing the nation's technological standing. It creates positive coverage in both domestic and international press at a moment when that coverage reaches unusually attentive political audiences. And it demonstrates — without requiring any government official to make the claim directly — that Chinese AI development is progressing despite export controls, despite Western competitive pressure, and despite the resource constraints that U.S. policy was designed to impose.
DeepSeek is a private company, not a state enterprise. But in China's economic system, the distinction between private and state interests in strategically significant technology sectors is more permeable than in Western market structures. The timing of V4's release is consistent with a company that is aware of its symbolic function in the national AI narrative, and is willing to coordinate its release cadence with that narrative's most important annual moment.
The "Sputnik moment" framing was first applied to DeepSeek after R1's January 2025 release. It is worth tracing the pattern across all three, because the pattern is more deliberate than any single release suggests.
Sputnik 1: DeepSeek R1 (January 2025). A reasoning model that matched OpenAI o1's performance at a fraction of the cost, released open-source. The Western response was a combination of denial (questioning benchmark reliability), partial acknowledgment (noting the cost efficiency), and accelerated investment (SoftBank's $500B Stargate announcement in the same news cycle). The stock market reaction was real: NVIDIA shed approximately $600B in market capitalization in a single day on concerns that cheaper AI would reduce demand for expensive compute.
Sputnik 2: DeepSeek V3 (late 2025). General-purpose capability benchmarks matching GPT-4o at significantly lower cost, again open-source. The Western response was faster acknowledgment and more aggressive competitive pricing moves. The framing shifted from "this is impossible" to "this is a cost problem, not a capability problem." Enterprise procurement teams began asking their AI vendors to justify their pricing against V3.
Sputnik 3: DeepSeek V4 (March 2026). Frontier-level capability (1T params, native multimodal, 1M context) at pricing 50x below GPT-5.2, open weights, optimized for export-controlled-circumventing hardware. The framing can no longer be "cost problem, not capability problem" because V4 is competing with the most capable models available in March 2026, not last year's frontier.
The escalation across three releases follows a consistent logic: each release addresses the specific objection that the Western AI establishment raised against the previous one. R1 was a reasoning model, not a general model. V3 was general but not multimodal. V4 is multimodal, context-extended, and priced to make the previous objections irrelevant.
The persistent Western analytical error about DeepSeek is attributing its cost advantage to factors that are either temporary or illegitimate: government subsidies, IP theft, distorted benchmarks, or the one-time benefits of engineering a smaller model. Each version of this critique has been progressively harder to sustain.
The more accurate account of DeepSeek's cost efficiency begins with MoE architecture. DeepSeek was an early and aggressive adopter of MoE for its flagship models — earlier and more successfully than OpenAI or Anthropic's flagship product lines. MoE's active-parameter efficiency is not a Chinese innovation, but exploiting it at production scale ahead of well-resourced Western competitors is a genuine engineering achievement.
The second factor is inference optimization. DeepSeek's engineering team has published technical reports documenting their CUDA-level kernel optimizations, memory management strategies, and hardware utilization improvements. These are not vague claims about efficiency; they are reproducible engineering artifacts. Western labs have similar expertise, but they have been building models at a scale that makes these optimizations a secondary priority after capability development.
The third factor is willingness to train on constrained hardware. NVIDIA export controls forced DeepSeek to optimize for the hardware it had, rather than scaling to the hardware available. Constraints tend to produce different engineering decisions than abundance. The Huawei Ascend optimization in V4 is partly a necessity-driven innovation — and necessity-driven innovations often produce better results than optionality-driven ones.
The pricing is the output of these three factors, not a strategic subsidy. An organization that genuinely costs $0.10 to serve a million tokens does not need government support to price at $0.10. The efficiency is real.
Benchmark claims in AI announcements require skepticism by default. DeepSeek's benchmark numbers on V4 at launch should be read with the following caveats.
DeepSeek's internal benchmarks show V4 matching or exceeding GPT-5.2 on MMLU (general knowledge), HumanEval (code generation), MATH (mathematical reasoning), and multimodal benchmarks including MMMU (Massive Multi-discipline Multimodal Understanding). The context window performance benchmarks show strong retention at 500K and 1M token lengths on needle-in-a-haystack evaluations.
What is not yet available: independent third-party replication of these benchmarks across diverse tasks, performance on the specific tasks that enterprise customers care about rather than academic benchmark suites, latency benchmarks under production load conditions, and comparative quality assessments on video understanding tasks where ground truth is harder to establish.
The pattern from R1 and V3 is that DeepSeek's self-reported benchmarks hold up to independent scrutiny better than initial Western skepticism suggested, but also reveal specific domains where V4 underperforms its claimed rankings. Long-form creative generation, culturally specific reasoning, and tasks requiring deep domain expertise in Western institutional contexts have historically been weaker for DeepSeek models than benchmark suites capture.
What this means for deployment decisions: evaluate V4 on your specific workload before making infrastructure commitments based on headline numbers. The headline numbers are not fabricated, but they are optimistic. Real-world performance on your task distribution may be 10–30% below benchmark performance, which is true of every major model and does not change the fundamental cost-efficiency argument.
The organizations for whom DeepSeek V4 is an immediate deployment consideration are those for whom cost is currently the binding constraint on AI adoption.
Document processing at scale. Legal, financial, and compliance teams processing large document volumes at GPT-5.2 pricing are spending 50x more than they need to for the same task quality. V4's 1M context window means document chunking becomes unnecessary for most enterprise document lengths. Self-host the weights for maximum cost reduction; use the API for lower-volume, variable workloads.
Video analysis pipelines. Computer vision teams that have deferred video-native AI because of cost can now build production pipelines on V4's native video support at API pricing that makes per-frame analysis economically viable.
Code review and refactoring. Codebases that previously required chunking due to context constraints can now be fed into V4 in their entirety, enabling codebase-aware refactoring recommendations that understand the full dependency graph.
Enterprises in cost-sensitive markets. Organizations in markets where AI cost has been prohibitive — emerging market enterprise, SMB segments, academic institutions — gain access to frontier-capable models at pricing that makes deployment viable.
The deployment cautions are real. Open weights mean your organization is responsible for security, compliance, and reliability rather than delegating those to a cloud provider. Self-hosting a 1T parameter MoE model requires substantial GPU infrastructure even with the active-parameter efficiency advantage. Data residency and regulatory compliance in your specific jurisdiction may limit the deployment options. And the geopolitical dimension — using a Chinese-developed model for sensitive enterprise workloads — is a legitimate risk assessment that different organizations will resolve differently.
DeepSeek V4 is not the end of this story. It is the third chapter of a story whose ending has not been written.
The Western response will come in two forms. First, pricing pressure: OpenAI and Anthropic will either reduce prices or introduce new lower-cost tiers that narrow the 50x differential. This has happened after each previous DeepSeek release. Second, capability escalation: the next generation of Western frontier models will attempt to establish a capability gap that V4 cannot match. This is the correct competitive response, and it is already underway — the investment levels at OpenAI, Anthropic, Google, and Meta in 2026 are unprecedented.
The open-source ecosystem response will also accelerate. V4's open weights will be fine-tuned, quantized, distilled, and deployed across hundreds of downstream applications within months of release. The open-source AI community, which has been building on Meta's Llama series, now has a 1T MoE model with frontier capability as its new foundation. The derivative models will be significant.
The geopolitical response is harder to predict. U.S. export controls will likely be reviewed in light of V4's Ascend optimization, but the design space for effective controls on AI development is genuinely difficult — it is much harder to restrict access to algorithmic innovations and training techniques than to restrict access to specific hardware.
The cost curve, which is the most important long-run trend, will continue downward. DeepSeek V4's pricing will not look shocking in 2027 because the frontier will have moved again. What V4 has accomplished is not locking in a permanent cost advantage but demonstrating that the cost curve can be broken faster than Western pricing structures assumed. That demonstration has already changed the AI infrastructure investment calculus for every enterprise budget conversation happening in Q2 2026.
DeepSeek V4 is a frontier-class large language model developed by DeepSeek, a Chinese AI lab. It uses a Mixture of Experts architecture with 1 trillion total parameters and 32 billion active parameters per token. It supports text, image, and video inputs natively, offers a 1 million token context window, and is available as open weights — meaning anyone can download and run the model weights without API dependency.
On API pricing, V4 is approximately 50x cheaper than GPT-5.2 at $0.10–$0.30 per million output tokens versus GPT-5.2's approximately $15 per million output tokens. On capability benchmarks, V4's self-reported numbers show competitive performance across MMLU, HumanEval, and MATH. Independent verification is ongoing. V4's 1M token context window exceeds GPT-5.2's standard 128K offering. GPT-5.2 benefits from a larger proprietary infrastructure, stronger Western regulatory compliance posture, and more mature tooling ecosystem.
DeepSeek releases model weights under a permissive license, which means the weights are publicly available for download and can be used for commercial purposes subject to the license terms. The training code, training data, and full technical specifications are not released — which is the standard definition of "open weights" rather than true open-source. The distinction matters: open weights allow deployment and fine-tuning but do not permit reproduction of the training process.
The risk assessment involves multiple dimensions. Technical: the model performs well on benchmarks and the open weights can be audited more thoroughly than a closed API. Regulatory: organizations in regulated industries must assess whether using a Chinese-developed model is compatible with their data handling obligations. Geopolitical: using a model developed by a Chinese company for sensitive enterprise workloads introduces supply chain and data sovereignty considerations that each organization must evaluate based on its specific regulatory environment and risk appetite. Many use cases — public data analysis, code assistance, document summarization — present lower risk than others.
Running a 1T MoE model with 32B active parameters requires substantial GPU infrastructure. In practice, efficient inference at reasonable throughput requires multiple high-memory GPUs — typically 8x A100 80GB or equivalent as a minimum for single-node deployment, with multi-node configurations needed for high-throughput production workloads. Quantized versions of the model (4-bit or 8-bit) reduce memory requirements substantially and will be available through the open-source community within weeks of the weights release.
DeepSeek has not publicly confirmed the timing was deliberate. The circumstantial evidence — a major capability announcement one day before China's most significant annual political event — is consistent with a company that understands its symbolic function in China's national AI narrative and times major announcements accordingly. The Two Sessions is the venue where China's government signals its technology priorities, and a frontier model release the day before ensures the conversation begins from a position of demonstrated capability.
It is their third. DeepSeek R1 in January 2025 was the first, demonstrating reasoning capability at open-source pricing. V3 in late 2025 was the second, extending that to general-purpose tasks. V4 is the third, adding native multimodal capability, 1M token context, and pricing 50x below the leading closed-source frontier. Each release has addressed the specific objection Western analysts raised against the previous one. Whether V4 constitutes a genuine capability parity or a cost-efficiency story that still trails on raw quality is a question that third-party evaluation over the next 60 days will answer more definitively than the launch numbers alone can.
DeepSeek denied US chipmakers early access to its V4 model while giving Huawei priority optimization time. Export control allegations add fuel.
Meta has quietly revealed its Llama 4 lineup: Mango (flagship multimodal) and Avocado (coding-specialized) models launching H1 2026, with open weights that could democratize AI access.
MiniMax released M2.5 — a 230 billion parameter mixture-of-experts model with 80.2% on SWE-Bench Verified, matching Claude Opus 4.6 speed at $0.29/M input tokens, trained across 200,000 real-world coding environments.