DeepSeek V4 Pro: 1.6 Trillion Parameters and the New Open-Source Frontier Race

What DeepSeek Actually Shipped on April 24

The preview release of DeepSeek V4 Pro on April 24, 2026 is the largest open-weight model ever made publicly available. Its 1.6 trillion total parameters exceed competitors including Kimi K 2.6 and M1, but the more architecturally significant figure is the 49 billion active parameters: V4 Pro uses a mixture-of-experts (MoE) design in which only a fraction of the total parameter count is engaged for any given inference. This is not a new technique — Mistral, Google, and others use MoE — but V4 Pro applies it at a scale that was previously the exclusive territory of closed, proprietary models.

The 1 million-token context window is similarly consequential. One million tokens covers approximately 750,000 words, which means V4 Pro can process entire legal contracts, codebases, or multi-year financial histories in a single context pass. Until V4 Pro, 1M-token context at open-weight was theoretical; it is now commercially available.

The pricing is where the disruption becomes most visible. At $0.145 per million input tokens and $3.48 per million output tokens, V4 Pro undercuts every closed-source model in its performance class. DeepSeek states the model outperforms its open-source peers across reasoning benchmarks and surpasses GPT-5.2 and Gemini 3.0 Pro on some tasks. On coding competition benchmarks, it is comparable to GPT-5.4. The acknowledged weaknesses — text-only support (no audio, video, or image capabilities) and a 3–6 month estimated lag behind frontier models on knowledge tests — are real but do not undercut the headline: for most enterprise use cases, V4 Pro’s capability-to-cost ratio is without precedent in the open-weight category.

What V4 Pro Changes in the Enterprise AI Calculus

The standard argument for using closed-source frontier models (GPT-5.x, Claude Opus 4.x, Gemini 3.x) has been capability: they are measurably better, and for high-stakes tasks the quality premium justifies the cost premium. V4 Pro complicates that argument in three specific ways.

First, it shifts the benchmark conversation. If V4 Pro is comparable to GPT-5.4 on coding competitions — a credible proxy for software engineering tasks — then the capability premium for closed models in software engineering workflows is no longer obvious. Enterprises running code generation, review, or refactoring at scale can now run honest cost-benefit comparisons rather than accepting closed-model pricing as the cost of quality.

Second, it resets pricing expectations. The $0.145/M input token price will become a reference point in every enterprise AI procurement conversation that happens through 2026. Vendors pricing above this level for comparable capability will face increasing pressure to justify the premium — whether through multimodal support, superior reasoning on specific benchmarks, or enterprise support SLAs that open-weight deployments do not provide.

Third, it accelerates the self-hosting conversation. V4 Pro is open-weight, which means enterprises with sufficient GPU infrastructure can run it on their own hardware. This matters most in regulated industries — financial services, healthcare, legal — where data sovereignty requirements make API-based AI deployments structurally difficult. A 1.6-trillion-parameter model that can be deployed on-premise, with institutional data never leaving the organisation’s perimeter, removes the primary structural barrier to AI adoption in the most AI-averse sectors.

What Enterprise AI Leaders Should Do About It

1. Benchmark V4 Pro against your closed-model spend before your next contract renewal

V4 Pro’s capability profile — strong on coding and reasoning, weaker on knowledge-intensive tasks, text-only, no multimodal — maps cleanly onto the most common enterprise AI use cases: document analysis, code generation, data extraction, and text summarisation. If your current closed-model usage is concentrated in these categories, V4 Pro is a legitimate substitution candidate. Run a structured benchmark: take 50-100 representative tasks from your current workload, run them through V4 Pro via API at $0.145/M input, compare output quality against your current model. Do this before your next vendor contract renewal, not after. The 3–6 month capability lag on knowledge tasks is real, but it is not uniform: many enterprise tasks are not knowledge-intensive in the benchmark sense, and V4 Pro’s performance gap on those tasks may be negligible.

2. Evaluate self-hosting V4 Pro for regulated data environments — the cost model changes dramatically at scale

The economics of self-hosted open-weight models versus closed-model API calls invert at scale. At low volume (under approximately 500 million tokens per month), API pricing is cost-effective and self-hosting is not justified by infrastructure cost. Above that threshold, self-hosted GPU compute typically becomes cheaper. For enterprises in financial services, healthcare, or public administration — where data sovereignty requirements create API-based AI barriers — V4 Pro’s open-weight availability means the cost model for compliant AI deployments has fundamentally changed. Run the calculation for your specific volume and data classification requirements before concluding that on-premise AI is economically infeasible.

3. Watch for the multimodal gap to close — and lock in API commitments now rather than at parity

V4 Pro’s text-only limitation is its most significant near-term weakness. DeepSeek’s previous model releases have followed a pattern of capability expansion over 6–12 months post-initial release. If V4 Pro follows that pattern, multimodal capabilities (image input at minimum) could arrive before the end of 2026. Enterprises that are waiting for open-weight multimodal at frontier pricing should treat April 2026 as the start of the countdown, not a reason to delay evaluation. Meanwhile, closed-model vendors whose multimodal capabilities represent their primary differentiator versus V4 Pro should be approached for longer-term contract commitments now — before open-weight multimodal parity forces a re-negotiation from a weaker leverage position.

The Bigger Picture: What Open-Weight at 1.6T Means for the Industry

DeepSeek V4 Pro is not an isolated product release. It is the latest evidence of a structural shift in the AI model market: the capability frontier is no longer the exclusive territory of closed-source systems backed by $100B+ capex cycles. V4 Pro was built by a relatively small team using architectural efficiency (MoE, reduced active parameter count) to reach performance levels that closed-source labs have spent billions to achieve.

This has implications that extend beyond pricing. Open-weight models create the conditions for genuine localisation: organisations can fine-tune V4 Pro on their own domain data, producing models that reflect their specific terminology, workflows, and data characteristics without dependency on closed-model providers. They enable capability preservation: if a closed-model vendor changes pricing, deprecates a model, or becomes geographically restricted, an open-weight alternative provides continuity. They also raise the bar for what “proprietary advantage” means in AI — a closed model that costs $3.48/M output tokens needs to be meaningfully better on the tasks that matter to a specific enterprise, not just better on aggregate benchmarks.

The text-only limitation and the 3–6 month knowledge benchmark lag are the honest constraints on V4 Pro as of April 2026. Neither is permanent. The open-source frontier race that V4 Pro accelerates will look materially different by Q4 2026. Enterprise AI procurement decisions made now should be positioned for that trajectory, not anchored to the closed-model status quo of Q1 2026.

Frequently Asked Questions

Is DeepSeek V4 Pro truly open-source, or is it open-weight?

V4 Pro is open-weight, not fully open-source. The distinction matters: open-weight means the model weights are publicly available for download and deployment, but DeepSeek has not published the training code, training data, or full methodology documentation. This is the industry standard for what most organisations call “open-source AI” — the model can be run, fine-tuned, and deployed without API dependency, but the full reproducibility of the training process is not guaranteed. For most enterprise use cases, the open-weight distinction is operationally irrelevant: what matters is the ability to run the model on your own hardware with your own data.

How does V4 Pro’s mixture-of-experts architecture affect deployment?

The MoE architecture means only 49 billion of V4 Pro’s 1.6 trillion parameters are active during any single inference pass. This reduces compute requirements compared to a dense model with equivalent total parameters — a 1.6T dense model would be effectively undeployable outside specialised clusters. In practice, V4 Pro’s GPU memory requirements are closer to a dense 70-100 billion parameter model, which is demanding but within reach of enterprise-grade multi-GPU setups. The trade-off is routing overhead: the system must determine which expert sub-networks to activate for each input, adding latency that dense models do not have.

What is the realistic timeline for DeepSeek V4 Pro to reach full multimodal capability?

DeepSeek’s release cadence suggests major capability additions arrive 6–12 months after the initial model launch. V3 (the predecessor to V4 Pro) received significant capability updates approximately 8 months post-release. Applying the same timeline to V4 Pro suggests multimodal input capability could arrive by Q4 2026 or Q1 2027. This is informed extrapolation, not a public commitment from DeepSeek — the company does not publish roadmaps. Enterprises that require multimodal AI capabilities now should not wait for V4 Pro; enterprises that can operate with text-only for 6–12 months gain V4 Pro’s substantial cost and context advantages in the interim.

—

Sources & Further Reading

• DeepSeek Previews New AI Model That Closes the Gap with Frontier Models — TechCrunch
• DeepSeek Unveils Newest Flagship, a Year After AI Breakthrough — Bloomberg
• Why DeepSeek’s V4 Matters — MIT Technology Review
• DeepSeek V4 LLM Preview: Open-Source AI Competition — CNBC
• China’s DeepSeek Releases New AI Model V4 — Euronews