DeepSeek V4 — Open-Weight Trillion-Parameter MoE at ~1/6th Frontier Cost
DeepSeek V4 (open-weight MIT): V4-Pro 1.6T/49B-active, V4-Flash 284B/13B-active, 1M context, ~$0.435/$0.87 per-Mtok — near-frontier capability (SWE-bench Verified ~80.6%, GPQA ~90–92%) at ~1/6th the cost of Opus 4.7 / GPT-5.5; lands inside an 8-day April-2026 frontier window.
DeepSeek V4 — Open-Weight Trillion-Parameter MoE at ~1/6th Frontier Cost
Fetch note: the primary VentureBeat URL returned HTTP 429. This body is written from the WebSearch summary plus a corroborating spec sheet (morphllm.com/deepseek-v4) and the Hugging Face model card. The source URL is preserved above. Benchmark figures are vendor- or third-party-reported, not peer-reviewed.
Core Thesis
DeepSeek V4 (released 24 April 2026, open-weight under MIT on Hugging Face) is the cost-disruption entry in the April-2026 frontier wave. Where Claude Opus 4.7 (Apr 16) and GPT-5.5 (Apr 23) are closed and premium-priced, V4 reaches near-state-of-the-art capability at ~1/6th the cost and ships open weights. It lands the same week as GPT-5.5 — the three releases compress the frontier into an ~8-day window.
Key Facts
Architecture (MoE family, two variants)
| Metric | V4-Pro | V4-Flash |
|---|---|---|
| Total parameters | 1.6 trillion | 284 billion |
| Active parameters / token | 49B | 13B |
| Context window | 1M tokens | 1M tokens |
| Max output | 384K tokens | 384K tokens |
- Hybrid-attention MoE design. In the 1M-token setting, V4-Pro runs single-token inference at ~27% of V3.2's FLOPs (Flash ~10%) and reduces KV-cache memory to ~10% (Pro) / ~7% (Flash) of V3.2 — the efficiency basis for the pricing.
Pricing (per 1M tokens)
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| V4-Pro | $0.435 | $0.003625 | $0.87 |
| V4-Flash | $0.14 | $0.0028 | $0.28 |
- Headline framing: ~1/6th the cost of Opus 4.7 / GPT-5.5 for comparable intelligence.
Benchmarks (vendor / third-party reported)
- SWE-bench Verified: V4-Pro-Max 80.6% — described as the highest open-weights entry, around the Gemini 3.1 Pro tier. (Some trackers cite higher V4-Pro SWE-bench Verified numbers ~91%; subset/harness differences apply.)
- GPQA (PhD science): ~90.1–91.8% (Diamond vs PhD-subset variation).
- SWE-bench Pro (harder split): ~55% for V4-Pro vs Opus 4.7 ~64% and GPT-5.5 ~59% — closed models still lead the hardest agentic-coding split.
- DeepSeek self-reported launch figure: LiveCodeBench Pass@1 ~93.5 (vendor-run, unverified).
Positioning in the wave
- Opus 4.7 (Apr 16) — coding precision + safety leadership ("Project Glasswing"); SWE-bench Verified reported ~87–94% across configs.
- GPT-5.5 "Spud" (Apr 23) — agentic versatility / knowledge work.
- DeepSeek V4 (Apr 24) — open-weight cost efficiency; closes most of the capability gap at a fraction of the price.
Significance
V4 is the clearest 2026 instance of the open-weight frontier compressing the closed-model lead to a thin, hardest-tasks-only margin while collapsing price. The "frontier" is now a tight cluster (Google, OpenAI, Anthropic, Meta + DeepSeek/Qwen/Kimi) where the differentiator is increasingly cost, openness, and agentic robustness on the hardest splits — not raw capability on saturated benchmarks.
Source: VentureBeat — DeepSeek V4 arrives with near-state-of-the-art intelligence at 1/6th the cost of Opus 4.7, GPT-5.5, 24 April 2026. Specs corroborated via morphllm.com/deepseek-v4 and the DeepSeek-V4-Pro Hugging Face model card.