April 24, 2026. DeepSeek released V4-Pro and V4-Flash — two new open-source models that arrive roughly a year after the R1 "Sputnik moment" and immediately stake a claim on coding benchmarks while pricing themselves far below Western frontier models.
Both are released under the MIT license with open weights. Both carry a 1 million token context window as standard. Both are available via API today.
The Two Models
DeepSeek-V4-Pro
The flagship. Positioned as a frontier-class model with competitive coding performance:
- 1.6 trillion total parameters / 49 billion active (Mixture of Experts)
- 865GB on Hugging Face
- Targets agentic coding, world knowledge, and STEM reasoning
- DeepSeek says it trails GPT-5.5 and Gemini 3.1 Pro by "approximately 3 to 6 months"
DeepSeek-V4-Flash
The fast and cheap alternative — not a stripped-down version:
- 284 billion total parameters / 13 billion active (Mixture of Experts)
- 160GB on Hugging Face
- Comparable reasoning to V4-Pro with faster response times
- Positioned as "the fast, efficient, and economical choice"
Both models support three reasoning effort modes: standard, think, and think-max (the equivalent of thinking on/off/extended).
Benchmark Results
V4-Pro
| Benchmark | V4-Pro | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|
| LiveCodeBench | 93.5% | 88.8% | — | 91.7% |
| Codeforces | 3206 | — | 3168 | — |
| SWE-bench Verified | 80.6% | 80.8% | — | — |
| Terminal-Bench 2.0 | 67.9% | 65.4% | 75.1% | — |
| HLE (reasoning) | 37.7% | 40.0% | 39.8% | 44.4% |
| HMMT 2026 math | 95.2% | 96.2% | 97.7% | — |
| SimpleQA-Verified | 57.9% | — | — | 75.6% |
Where V4-Pro leads: coding. LiveCodeBench 93.5% is the highest score among all evaluated models. Codeforces 3206 beats GPT-5.4's 3168.
Where it trails: general knowledge recall (SimpleQA 57.9% vs Gemini's 75.6%) and multidisciplinary reasoning (HLE 37.7% — below all three competitors).
SWE-bench at 80.6% is within 0.2 points of Claude Opus 4.6's 80.8% — and both are well below Claude Opus 4.7's 87.6% shipped last week.
V4-Flash
| Benchmark | V4-Flash | V4-Pro |
|---|
| SWE-bench Verified | 79.0% | 80.6% |
| LiveCodeBench | 91.6% | 93.5% |
| Terminal-Bench 2.0 | 56.9% | 67.9% |
| SimpleQA-Verified | 34.1% | 57.9% |
Flash sits 1–2 points behind Pro on coding benchmarks — significant parity given the massive cost difference.
Pricing: The Defining Story
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|
| V4-Flash | $0.28 / 1M | $0.028 / 1M | $0.28 / 1M |
| V4-Pro | $1.74 / 1M | $0.145 / 1M | $3.48 / 1M |
| GPT-5.4 | $2.50 / 1M | $0.25 / 1M | $15.00 / 1M |
| GPT-5.5 | $5.00 / 1M | $0.50 / 1M | $30.00 / 1M |
| Claude Opus 4.6 | $5.00 / 1M | — | $25.00 / 1M |
Key comparisons:
- V4-Flash undercuts GPT-5.4 Nano ($0.20/$1.25 input/output) on input price — comparable output cost
- V4-Pro is 7x cheaper than Claude Opus 4.6 on input, while matching it on SWE-bench
- V4-Pro output is 4.3x cheaper than GPT-5.4 — and nearly 9x cheaper than GPT-5.5
- V4-Flash output at $0.28/1M is 89x cheaper than Claude Opus 4.6
This pricing, combined with the MIT license, is what makes DeepSeek V4 disruptive. The model isn't quite at the frontier — but it's close enough, at a price that changes the math for production deployments.
Architecture: Three Core Innovations
1. Hybrid Attention — CSA + DSA
V4 introduces "Novel Attention" combining Compressed Sparse Attention (CSA) with DeepSeek Sparse Attention (DSA), featuring token-wise compression. The result vs V3.2:
- V4-Pro: only 27% of single-token FLOPs, 10% of KV cache size
- V4-Flash: only 10% of single-token FLOPs, 7% of KV cache size
The 1M context window becomes practical (not just theoretical) because the KV cache costs are dramatically reduced.
2. Manifold-Constrained Hyper-Connections (mHC)
Training a 1.6T parameter model at this scale requires solving gradient stability problems. mHC reduces signal amplification from 3,000x down to 1.6x — enabling stable training runs that wouldn't converge at this parameter count otherwise.
3. Muon Optimizer
Both models trained with Muon instead of AdamW, delivering faster convergence and greater training stability:
- V4-Pro trained on 33 trillion tokens
- V4-Flash trained on 32 trillion tokens
Access and Integration
Chat Interface
API
Available immediately at api.deepseek.com. Two compatibility modes:
# OpenAI-compatible
from openai import OpenAI
client = OpenAI(api_key="your-key", base_url="https://api.deepseek.com")
response = client.chat.completions.create(
model="deepseek-v4-pro", # or deepseek-v4-flash
messages=[{"role": "user", "content": "..."}],
)
# Anthropic-compatible
import anthropic
client = anthropic.Anthropic(api_key="your-key", base_url="https://api.deepseek.com")
message = client.messages.create(
model="deepseek-v4-pro",
max_tokens=1024,
messages=[{"role": "user", "content": "..."}],
)
Compatible with Claude Code, OpenClaw, and OpenCode agentic systems as drop-in replacements.
Open Weights (Hugging Face)
Both models available for self-hosting under MIT license:
- V4-Pro: 865GB — requires H100/H200 infrastructure
- V4-Flash: 160GB — more accessible for teams with A100 clusters
Note: No Jinja-format chat template at launch — use the Python encoding scripts from the Hugging Face repository.
Legacy Model Retirement
Models deepseek-chat and deepseek-reasoner retire on July 24, 2026. Update your model parameter to deepseek-v4-pro or deepseek-v4-flash before that date.
Limitations
Honest assessment of the gaps:
- General knowledge recall is weak. SimpleQA-Verified at 57.9% trails Gemini 3.1 Pro's 75.6% significantly.
- Multidisciplinary reasoning trails. HLE at 37.7% is below GPT-5.4 (39.8%), Claude Opus 4.6 (40.0%), and Gemini (44.4%).
- Preview status. Further post-training refinements are expected — these are not final-quality releases.
- China-based API infrastructure. Data sovereignty is a real consideration for regulated industries.
- Self-hosting V4-Pro is resource-heavy. 865GB and H100/H200 requirements put it out of reach for most teams.
- No Jinja chat template. Minor but requires workaround at launch.
How It Compares at a Glance
| V4-Pro | V4-Flash | GPT-5.5 | Claude Opus 4.7 |
|---|
| Parameters (active) | 49B | 13B | Undisclosed | Undisclosed |
| Context window | 1M | 1M | Standard | 200K |
| License | MIT (open) | MIT (open) | Closed | Closed |
| Input price | $1.74/M | $0.28/M | $5.00/M | $5.00/M |
| Output price | $3.48/M | $0.28/M | $30.00/M | $25.00/M |
| LiveCodeBench | 93.5% | 91.6% | — | — |
| SWE-bench | 80.6% | 79.0% | — | 87.6% |
| HLE | 37.7% | — | — | 54.7% |
| API now | Yes | Yes | Coming soon | Yes |
Bottom Line
DeepSeek V4 isn't frontier-of-frontier. By DeepSeek's own admission, it trails GPT-5.5 and Gemini 3.1 Pro by 3–6 months in capability. On knowledge recall and general reasoning, the gap is real.
But on coding — the benchmark that matters most for developer and agentic use cases — V4-Pro leads the pack on LiveCodeBench and matches Claude on SWE-bench at a fraction of the price. V4-Flash delivers near-Pro coding performance at costs that make it viable as a default inference model in production.
For teams where budget matters and the primary use case is code, V4 is now the most interesting open-weight option available. For teams that need strong factual grounding or broad reasoning depth, the gaps in SimpleQA and HLE are worth weighing carefully.
API access: api.deepseek.com — available now.
Chat: chat.deepseek.com — Expert Mode (Pro) or Instant Mode (Flash).
Open weights: Hugging Face — MIT license, self-hosting available.
Sources:
