1. DeepSeek V4: 1.6 Trillion MoE, Fully Open Under MIT License
On July 20, 2026, DeepSeek released V4 GA — the company's most capable model generation to date. V4 ships in two variants: V4-Pro and V4-Flash, both released under the MIT open-source license.
| Metric | DeepSeek V4-Pro | DeepSeek V4-Flash |
|---|---|---|
| Architecture | MoE, 1.6T total / 49B active | 284B total / 13B active |
| Context window | 1 million tokens | 1 million tokens |
| License | MIT | MIT |
| Pricing (output) | Custom | $0.28 / million tokens |
| Codeforces Rating | 3,206 | — |
V4 marks DeepSeek's entry into structured, time-based pricing: the company introduced peak-valley time-of-use billing, with prices doubling during designated weekday peak hours. Legacy endpoints deepseek-chat and deepseek-reasoner will be retired on July 24.
This release completes an extraordinary seven-day window: Kimi K3 (2.8T, Frontend Code Arena #1, July 16), Qwen3.8 (2.4T, confirmed for release, July 19), and now DeepSeek V4 (1.6T, Codeforces #1 among open models, July 20). Three independently developed Chinese frontier open-source models, all achieving global top-tier performance, all released in the same week.
2. Codeforces Rating 3,206: Coding Capability Beyond GPT-5.4
The headline performance metric for DeepSeek V4-Pro is its Codeforces competitive programming rating of 3,206 — a score that exceeds GPT-5.4 and establishes V4-Pro as the strongest open-source model on algorithmic reasoning benchmarks.
| Benchmark | DeepSeek V4-Pro |
|---|---|
| Codeforces Rating | 3,206 (above GPT-5.4) |
| V4-Flash vs V3.2 | Flash outperforms previous flagship on most benchmarks |
| Competitive programming | Strongest open-source model |
Codeforces is a particularly meaningful benchmark for real-world coding: it measures algorithmic problem-solving under time constraints — the type of reasoning that translates directly to software engineering, data analysis, and automated workflow construction. For brands building AI-powered marketing tools, this capability means models that can genuinely reason through complex automation pipelines rather than pattern-match through templates.
V4-Flash, despite being the smaller variant at 284B parameters (13B active), outperforms the previous flagship V3.2 on most evaluation tasks — a signal that DeepSeek's training methodology improvements are more significant than raw parameter scaling.
3. $0.28/M Tokens and Peak-Valley Pricing: A New Pricing Paradigm
DeepSeek V4 introduces the industry's first peak-valley time-of-use pricing for an AI model API:
| Mechanism | Detail |
|---|---|
| Base price (Flash, output) | $0.28 / million tokens |
| Peak-hour multiplier | 2× during designated weekday windows |
| Off-peak discount | Base rate applies during nights and weekends |
This pricing model has significant implications for brand marketing operations:
1. Batch content generation moves to off-peak hours. Bulk copy generation, product description updates, and multilingual translation pipelines can be scheduled for off-peak windows where the $0.28 base rate applies — a direct 50% cost reduction compared to peak-hour rates.
2. Real-time interactive applications pay a premium. Customer-facing chatbots and live marketing tools that must respond during business hours face the 2× multiplier. The $0.56 peak rate remains competitive (still ~90% cheaper than GPT-5.6 Sol), but the price signal explicitly differentiates between batch and real-time workloads.
3. The MIT license eliminates vendor lock-in. Unlike Kimi K3's open-weights-by-July-27 timeline or Qwen3.8's planned release, DeepSeek V4 ships with full MIT-licensed weights and code available immediately. Brands can self-host V4 with zero licensing fees, zero API dependency, and full data sovereignty — the strongest possible compliance posture for regulated industries.
4. The Dual-Model Strategy: Pro for Reasoning, Flash for Scale
DeepSeek's V4 lineup reflects a deliberate dual-model strategy:
| Use case | Model | Why |
|---|---|---|
| Complex reasoning, competitive coding, research | V4-Pro (49B active) | Maximum intelligence per inference |
| High-volume production, content generation, agent pipelines | V4-Flash (13B active) | Near-frontier quality at 5× lower cost |
| Nighttime batch jobs, scheduled workflows | V4-Flash (off-peak) | $0.28 base rate, 50% discount vs peak |
This mirrors the architectural strategy pioneered by DeepSeek with V3: separate reasoning and high-throughput variants that share the same training foundation but optimize for different deployment profiles. The key difference in V4 is that both variants now converge on near-identical quality envelopes — Flash isn't a "dumb" version, it's a "fast" version.
For brand marketers, the practical workflow is straightforward: use V4-Pro for strategy (campaign architecture, competitive analysis, creative brief generation) and V4-Flash for execution (copy variants, A/B test generation, multilingual translation, landing page code).
5. China's Open-Source Model Week: Kimi K3 → Qwen3.8 → DeepSeek V4
The July 14–20 window will be remembered as a watershed moment in AI history. Three independently developed Chinese frontier models — each achieving global top-tier performance — released within the same seven days:
| Date | Model | Distinction | Openness |
|---|---|---|---|
| July 16 | Kimi K3 (2.8T) | Frontend Code Arena #1, world's largest open-source | Open weights July 27 |
| July 19 | Qwen3.8 (2.4T) | Second only to Fable 5,阿里云生态 | Confirmed open-source |
| July 20 | DeepSeek V4 (1.6T) | Codeforces 3206 > GPT-5.4, peak-valley pricing | MIT, available now |
This is not a coincidence of timing. It reflects three independent companies independently reaching the frontier of model quality — and all choosing to release openly. The structural implications are significant: if an AI capability can be deployed via an open-source, self-hosted model, the commercial API market for that capability faces immediate price compression.
6. What This Means for Brands
For TMG's audience of cross-border brand marketers, DeepSeek V4's release crystallizes three strategic implications:
1. The cost floor for AI-powered marketing has collapsed. Between DeepSeek V4-Flash at $0.28/M tokens (MIT) and Kimi K3 at $15/M tokens (open weights), brands now have options spanning a 50× price range — all at frontier or near-frontier quality. The economic logic of "we can't afford AI at scale" no longer holds.
2. Self-hosted deployment is now the compliance default, not the aspirational goal. MIT licensing + 13B active parameters (Flash) means a brand can deploy a production-grade AI marketing engine on a single on-premises server, with zero data leaving the enterprise perimeter. This directly addresses the July 15 AI-personification ban's data sovereignty requirements.
3. Peak-valley pricing introduces operational discipline to AI marketing. Brands that build batch content pipelines (nightly multilingual blog translation, weekly product description generation, scheduled A/B test variant creation) can achieve 50% lower inference costs through simple scheduling. AI cost management is now an operational skill, not just a procurement decision.
Key Takeaways
- DeepSeek V4 released July 20 — Pro (1.6T MoE, 49B active) and Flash (284B, 13B active), both MIT open-source
- V4-Pro Codeforces Rating 3,206 — exceeds GPT-5.4, strongest open-source model on algorithmic reasoning
- V4-Flash at $0.28/M output tokens — among the cheapest frontier-grade models available
- Industry-first peak-valley pricing — 2× multiplier during weekday peak hours, base rate at nights/weekends
- Legacy endpoints retired July 24 — full migration deadline
- China's open-source model week — Kimi K3 (7/16), Qwen3.8 (7/19), DeepSeek V4 (7/20) — three frontier models, seven days, all open