DeepSeek/deepseek-v4-flash

Supported Functionality

Item	Specification
Input	Text
Output	Text
Context	1,048,576 tokens (1M)
Max Output	Shares the 1M window (Think Max mode recommends ≥384K output budget)
Vision	✗ Not Supported (entire V4 family is text-only)
Function Calling	✓ Supported (dual-compatible with OpenAI ChatCompletions and Anthropic Messages APIs)

Description

DeepSeek-V4-Flash is the efficiency-tier flagship in DeepSeek's V4 family, released April 24, 2026 alongside V4-Pro. It is a Mixture-of-Experts model with 284B total parameters and 13B activated per token, separately pretrained on 32T+ tokens (not a distillation of Pro), and shipped as open weights on Hugging Face under the MIT license. The model is built around aggressive long-context efficiency: at 1M context, KV cache is just 10% of V3.2's and per-token FLOPs drop sharply.

Flash's core positioning is "the open-source production workhorse with the best price-performance." It scores 47 on the Artificial Analysis Intelligence Index, matches V4-Pro on simple agent tasks, yet costs roughly 1/12 of Pro per output token. It supports dual Thinking / Non-Thinking modes plus three reasoning-effort tiers (high / xhigh / max), letting one model span fast chat and deep reasoning. On the official DeepSeek API, the legacy deepseek-chat and deepseek-reasoner endpoints now route to V4-Flash non-thinking / thinking by default.

Key Capabilities

• Efficient Sparse MoE Inference: 284B total / 13B active gives 13B-dense-model inference cost with intelligence well above that tier.
• Production-Ready 1M Context: Hybrid sparse attention cuts KV cache to 10% of V3.2 and per-token FLOPs to ~27% — whole-repo code and long-document workloads are economically viable.
• Dual-Mode + Three Reasoning Tiers: Thinking / Non-Thinking modes plus high / xhigh / max effort cover everything from low-latency chat to deep reasoning.
• Reasoning Close to Pro: Flash@Max approaches V4-Pro@High on most reasoning tasks; gap shows mainly on pure-knowledge queries and the hardest agent loops.
• Open-Source SOTA Agentic Coding: V4 leads open-source agentic coding benchmarks, and Flash's price makes it the default self-hosted Cursor/Copilot replacement.
• Dual API Compatibility: OpenAI ChatCompletions and Anthropic Messages APIs both supported — migration is effectively a base_url/model swap.
• Full Open Weights under MIT: 160 GB (FP8) on Hugging Face; deployable on a single A100 80GB or dual RTX 4090.

Technical Strengths

Feature	Benefit
Sparse MoE with 13B active	Pro-tier intelligence at roughly the inference cost of a 13B dense model
Hybrid Sparse Attention	KV cache at 10% of V3.2 at 1M context — long-context RAG and document QA become cheap
Independent 32T+ token pretrain	Not a distillation of V4-Pro; same architectural family, higher capability ceiling than a derivative
Thinking / Non-Thinking dual mode	One model serves both low-latency support and deep reasoning — no model-switching overhead
Disruptive pricing ($0.14 / $0.28)	$0.14 in / $0.28 out per 1M tokens beats even GPT-5.4 Nano — a TCO killer at scale
Drop-in OpenAI / Anthropic compatibility	Swap base_url and model name; existing client code keeps working

Capability Ratings

Dimension	Rating	Notes
Reasoning	Excellent	AA Intelligence Index 47; Flash@Max approaches V4-Pro@High — top tier among open models
Coding	Excellent	SOTA on open-source agentic coding benchmarks; one A100 saturates a 5-developer team's needs
Creative Writing	Strong	Solid CN/EN writing; lags V4-Pro and closed frontier models on long-form creative tasks
Multimodal	Moderate	Text-only input/output; no native vision or audio support
Response Speed	Very Fast	13B active params + sparse attention give very high throughput and low latency
Context Window	Huge	1M tokens with dramatically lower long-context cost than dense peers

Use Cases

• High-Volume RAG & Customer Support: 13B-active inference plus 1M context is the cost-optimal choice for large knowledge-base QA and support systems.
• Self-Hosted Coding Assistants: MIT license and single-A100 deployment make it a viable team-level Cursor/Copilot replacement saturating ~5 developers per node.
• Long-Document & Whole-Repo Analysis: 1M context ingests entire books, repositories, or hour-long transcripts; sparse attention keeps cost low.
• Smooth API Migration: Existing OpenAI- or Anthropic-protocol apps can switch by changing only base_url and model name — no client rewrite needed.
• Reasoning-Heavy Workflows: Math competitions, complex planning, multi-step agent tasks reach near-Pro quality via xhigh / max thinking mode.
• Domain Fine-Tuning Base: Full MIT weights and clean MoE architecture make it an ideal base for SFT/DPO in medicine, law, and finance.
• Budget-Sensitive Developers & Startups: At $0.14 / $0.28 per 1M tokens, even small teams can run frontier-class intelligence at production scale.