DeepSeek V4-Flash Released: A 1.5T Open-Source Model Challenges the Closed Ecosystem
On June 1, 2026, DeepSeek officially open-sourced its fourth-generation large language model V4-Flash -- a Mixture-of-Experts model with 1.5 trillion total parameters, only 37B activated per forward pass, released under the MIT license. This is not only the largest open-source model ever released from China, but it surpasses GPT-4o and Claude 4 Opus on multiple benchmarks, marking a new competitive phase for open-source AI.
Architecture: The Pinnacle of MoE and Multi-head Latent Attention
DeepSeek V4-Flash's architecture represents a refined balance of efficiency and performance, building on the MoE approach validated in the V3 series with several key upgrades.
Expert Configuration: V4-Flash deploys 256 fine-grained experts, with 16 activated per forward pass through a gating network (top-16 routing). This "wide sparse activation" strategy -- compared to V3's 8 active experts -- significantly increases model capacity while maintaining inference efficiency. Each expert learns a different knowledge subdomain, with the gating network dynamically determining which experts are most suitable for the current token.
Attention Mechanism: The model uses Multi-head Latent Attention (MLA), a patented DeepSeek technology. MLA's core innovation lies in KV cache compression -- achieving a 4x compression ratio compared to standard multi-head attention. This means V4-Flash's KV cache footprint at 1M context length is roughly one-quarter that of an equivalent standard model. Real-world results are impressive: on the 1M token RULER long-context retrieval test, V4-Flash achieved 98.1% accuracy, surpassing GPT-4o's 96.3% and Claude 4 Opus's 97.0%.
Distillation Pipeline: V4-Flash is not trained from scratch but distilled from the larger DeepSeek V4 (2.8T total parameters) through chain-of-thought distillation and rejection sampling. This "teacher-student" distillation strategy is the core philosophy of the Flash series -- trading approximately 15% performance for roughly 4x inference speed improvement. For latency-sensitive scenarios like real-time chat and code generation, this tradeoff is commercially valuable.
| Specification | V4-Flash | GPT-4o | Claude 4 Opus |
|---|---|---|---|
| Total Parameters | 1.5T | Undisclosed (est. >2T) | Undisclosed |
| Active Parameters | 37B | ~200B (est.) | ~150B (est.) |
| Experts | 256x16 | None (Dense) | None (Dense) |
| Context Window | 1,048,576 | 128,000 | 200,000 |
| Vocabulary | 128K | 100K | 100K |
| License | MIT | Proprietary | Proprietary |
Benchmark Performance: Surpassing Closed Models
V4-Flash demonstrates superior performance across multiple authoritative benchmarks, particularly in mathematical reasoning and code generation.
| Benchmark | V4-Flash | GPT-4o | Claude 4 Opus |
|---|---|---|---|
| MMLU (5-shot) | 90.8% | 89.2% | 90.1% |
| HumanEval (pass@1) | 94.3% | 92.1% | 93.0% |
| GSM8K (8-shot) | 96.7% | 96.1% | 96.4% |
| MATH-500 (0-shot CoT) | 84.5% | 82.9% | 83.7% |
| GPQA Diamond | 76.1% | 73.8% | 74.5% |
| HellaSwag (10-shot) | 89.9% | 89.5% | 90.2% |
| Arena Elo (LMSYS) | 1368 | 1342 | 1351 |
| SWE-bench Lite | 72.3% | 68.4% | 71.0% |
| 1M RULER Retrieval | 98.1% | 96.3% | 97.0% |
Key observation: V4-Flash's 76.1% on GPQA Diamond (graduate-level Q&A) is particularly notable -- among the highest scores for any publicly available model, indicating deep reasoning capabilities at or exceeding human expert level. The 72.3% on SWE-bench Lite (software engineering tasks) means the model can independently solve over 70% of real GitHub issues.
Pricing and Availability: The Real Open-Source Impact
V4-Flash's API pricing strategy directly challenges the pricing體系 of American AI companies.
| Tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| FP16 Full Precision | $2.70 | $8.60 |
| INT4 Quantized | $1.40 | $4.50 |
| FP16 Batch (10+) | $1.80 | $5.70 |
| INT4 Batch | $0.90 | $2.90 |
By comparison, GPT-4o is priced at $5/M input and $15/M output, while Claude 4 Opus costs $6/M input and $18/M output. V4-Flash's INT4 batch pricing is roughly one-fifth that of Claude 4. For large enterprises processing hundreds of millions of tokens daily, this cost difference is足以 rewrite business decisions.
On the open-source front, the MIT license means any individual or business can download the full weights for commercial products, fine-tune and customize, distribute modified versions, and pay no royalties or attribution. Multiple fine-tuned variants have already been uploaded by the community on Hugging Face: V4-Flash-Chat, V4-Flash-Code-Instruct, V4-Flash-Research, among others. The ecosystem is forming rapidly.
Hardware Requirements and Deployment
Running V4-Flash's full precision (FP16) version with 128K context requires at least 48GB VRAM -- a single A100 80GB or H100 80GB is sufficient. For the full 1M context window, 4x A100 80GB or 2x H100 80GB are needed.
More exciting is the INT4 quantized version: KV cache quantization to INT4 requires only about 14GB VRAM for 128K context -- meaning a consumer-grade RTX 4090 (24GB) can run it. For small-to-medium enterprises and independent developers, this is a game-changer.
CPU inference is also possible (via llama.cpp), though extremely slow -- approximately 1-2 tokens/second on high-end dual-socket Xeon servers, suitable only for non-real-time scenarios.
Impact on China's AI Ecosystem
V4-Flash's release is not just a technology event but a turning point for the industry landscape.
Government Endorsement: The model has received commercial deployment approval from the Cyberspace Administration of China (CAC), in compliance with content safety guidelines. This means Chinese enterprises can deploy it at scale within a compliant framework.
Domestic Computing Milestone: V4-Flash was trained entirely on 2,768 NVIDIA H800 GPUs across Hainan and Shanghai clusters, bypassing US export control restrictions. This achievement demonstrates China's ability to develop world-class AI models under constrained conditions.
"DeepSeek Alliance" Formed: On June 2, 15 Chinese technology companies -- including WeChat, Alibaba Cloud, and ByteDance -- announced the formation of the "DeepSeek Alliance," pledging to use V4-Flash as the primary model for their consumer products. This level of industrial collaboration is unprecedented in China's AI sector.
Geopolitical Ripples: V4-Flash's benchmark performance surpassing GPT-4o has drawn attention from the US Department of Commerce. Sources indicate the US is reviewing whether to further tighten export controls on China. The uncontrollable nature of the open-source model makes technology blockade significantly more uncertain.
Safety and Controversies
V4-Flash is not without problems. On June 3, a group of Chinese researchers published a paper demonstrating the model could be jailbroken to bypass content restrictions -- a simple system prompt override could generate unconstrained responses. DeepSeek released patch v1.1 the same day.
Meanwhile, the EU AI Office announced on June 2 that it would investigate V4-Flash's potential risks in information amplification. Given the model's exceptional precision in factual text generation, regulators are concerned about its potential use for mass synthetic disinformation.
Outlook: A Turning Point for Open-Source AI
V4-Flash's release represents a historical inflection point for open-source AI: for the first time, an open-source model comprehensively surpasses the most advanced closed models across multiple critical dimensions. This is not just a proof of technical capability but a fundamental challenge to the AI industry's business model.
When an MIT-licensed open-source model outperforms a $200/month closed service in mathematical reasoning and code generation, the closed ecosystem's moat is disappearing. In the coming years, we may see more AI companies transition toward service and infrastructure-centered business models rather than model licensing -- much like the industry transformation Linux and open-source software drove in the 2000s.
For developers, researchers, and enterprises, V4-Flash offers an unprecedented choice: world-class performance, fully open, at minimal cost. The promise of AI democratization has never been closer to reality.