DeepSeek-V4-Flash

// Overview

DeepSeek-V4-Flash is the smaller, faster, and cheaper of the two models in DeepSeek's V4 series, released as a preview on 24 April 2026 alongside the larger DeepSeek-V4-Pro. It is a Mixture-of-Experts language model with 284 billion total parameters, of which 13 billion are activated per token, and it natively supports a one-million-token context window with up to 384K tokens of output.

Like the Pro tier, DeepSeek-V4-Flash uses a Hybrid Attention Architecture that combines Compressed Sparse Attention and Heavily Compressed Attention. DeepSeek reports that in the 1M-token setting it needs only about 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2, which is what makes such a long context affordable. The model was pre-trained on more than 32 trillion tokens, runs in mixed FP4/FP8 precision (FP4 for MoE experts), and exposes three reasoning-effort modes: Non-think, Think High, and Think Max. DeepSeek says Flash in its Max mode approaches the reasoning quality of V4-Pro when given a larger thinking budget.

All weights are published on Hugging Face under the MIT License, so DeepSeek-V4-Flash is free for commercial use, fine-tuning, and self-hosting via vLLM, SGLang, or quantized GGUF builds in Ollama and LM Studio. On the DeepSeek API the legacy deepseek-chat and deepseek-reasoner aliases were retired on 2026-07-24 after having routed to Flash's non-thinking and thinking modes, and DeepSeek positions Flash as a low-cost agentic and coding workhorse at $0.14 per million input tokens.

On 2026-07-31 the `deepseek-v4-flash` API entered official public beta on a re-post-trained checkpoint DeepSeek released as DeepSeek-V4-Flash-0731. The checkpoint keeps the same architecture and parameter count as the April preview but reports substantially stronger agent scores — Terminal-Bench 2.1 82.7, DSBench-FullStack 68.7, Cybergym 76.7, Toolathlon Verified 70.3, DeepSWE 54.4, and NL2Repo 54.2, tested with the DeepSeek Harness minimal mode at max effort. The official V4-Flash natively supports the Responses API format and is specifically adapted for OpenAI Codex; DeepSeek says the official release of V4-Pro will follow.

// Benchmarks

1. MMLU-Pro (Max)86.2%
2. GPQA Diamond (Max)88.1%
3. HMMT 2026 Feb (Max)94.8%
4. LiveCodeBench (Max)91.6%
5. SWE Verified (Max)79%
6. SimpleQA-Verified (Max)34.1%
7. MRCR 1M (Max)78.7%
8. Terminal-Bench 2.1 (V4-Flash-0731)82.7%
9. NL2Repo (V4-Flash-0731)54.2%
10. Cybergym (V4-Flash-0731)76.7%
11. DeepSWE (V4-Flash-0731)54.4%
12. Toolathlon Verified (V4-Flash-0731)70.3%
13. DSBench-FullStack (V4-Flash-0731)68.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Pricing source ↗

// Strengths

• Open MIT-licensed weights, free for commercial use, fine-tuning, and local deployment via vLLM, SGLang, Ollama, or LM Studio
• Native one-million-token context with efficient hybrid attention (roughly 27% of V3.2's per-token FLOPs and 10% of its KV cache at 1M)
• Selectable reasoning effort (Non-think / Think High / Think Max) to trade latency for depth
• Very low API pricing — $0.14 per 1M input tokens and $0.28 per 1M output, about 12x cheaper than V4-Pro, with aggressive cache-hit discounts
• Strong coding and reasoning scores for its size, approaching V4-Pro in the Max mode (LiveCodeBench 91.6, SWE Verified 79.0)

// Best for

• High-volume agentic and coding workflows where cost and latency matter more than absolute peak quality
• Long-document and large-codebase analysis that needs the full 1M-token window on a budget
• Self-hosted / on-prem deployment where open MIT weights and a smaller 13B-active footprint are required
• A drop-in replacement for the retiring deepseek-chat and deepseek-reasoner endpoints via its non-thinking and thinking modes

// How to access

// DeepSeek V4 — every version

The full lineage of the DeepSeek V4 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.