DeepSeek V4 Flash: Everything You Need to Know About the 284B-Parameter Open-Source Model

DeepSeek has been on a publishing tear. On July 31, 2026, the company released the production version of DeepSeek-V4-Flash-0731 under the MIT license — an open-source, 284-billion parameter mixture-of-experts (MoE) model that rivals the performance of its much larger sibling, V4 Pro, on agentic and coding benchmarks . At the same time, DeepSeek is inviting open-source developers to beta test DeepSeek Harness, an agent execution framework that promises to turn LLMs into autonomous task-completing agents . Here's everything you need to know about these developments and what they mean for developers, enterprises, and the AI landscape.

What Is DeepSeek V4 Flash?

DeepSeek V4 Flash is the efficiency tier of the DeepSeek V4 family, designed for high-volume, latency-sensitive workloads. It uses a MoE architecture with 284 billion total parameters, of which only 13 billion are active per token . This means the model has access to a vast pool of specialized expert knowledge while keeping inference fast and cost-efficient.

Property	Value
Total parameters	284B (304B with DSpark draft module)
Active parameters	13B per token
Architecture	Mixture-of-Experts (MoE)
Context window	1 million tokens
Max output	384K tokens
Precision	FP4 + FP8 mixed
License	MIT
Download size	~160 GB

Sources:

MIT License and Open Weights

Both V4 Flash and its larger sibling V4 Pro are released under the permissive MIT license, which permits unlimited commercial use, modification, and distribution without royalties . The weights are available on Hugging Face and can be self-hosted on private infrastructure. DeepSeek V4 is the only open-weight model lineup in its tier with a 1M-token context window available under a permissive MIT license .

The Attention Breakthrough: CSA and HCA

The V4 Flash model achieves its massive 1-million-token context window through a hybrid attention architecture that combines two novel mechanisms:

• Compressed Sparse Attention (CSA): Compresses key-value entries by 4× along the sequence dimension using softmax-gated pooling. A "Lightning Indexer" running in FP4 then selects the top-512 compressed blocks per query, drastically reducing the attention computation .
• Heavily Compressed Attention (HCA): Provides a second layer of compression at a 128:1 ratio for extreme long-context scenarios. This path is dense (no sparse selection) and handles the bulk of historical context .

The model alternates between CSA and HCA layers: even layers from layer 2 onward use CSA (4:1 compression), while odd layers from layer 3 onward use HCA (128:1 compression). A sliding-window branch over the last 128 raw tokens is always present for recency .

FP4/FP8 Mixed Precision

The model checkpoint uses mixed-precision quantization to shrink memory footprint:

• MoE expert weights are stored in FP4 (MXFP4 format), packed two values per int8 byte .
• All other parameters — attention, normalization, router, and shared experts — remain in FP8 .
• An NVFP4 variant is also available from NVIDIA (nvidia/DeepSeek-V4-Flash-NVFP4) .

At FP8, the full 284B model weighs approximately 284 GB for weights alone. For self-hosting, the recommended minimum configuration for full 1M context is 4x NVIDIA H200 SXM5 (564 GB total VRAM) .

Configurable Reasoning Effort

DeepSeek V4 Flash exposes a configurable reasoning_effort parameter that lets developers choose the model's inference mode:

• Non-think — Fast, intuitive responses for routine tasks and low-risk decisions .
• Think High — Deliberate chain-of-thought reasoning for complex problem-solving .
• Think Max — Maximum reasoning effort, using a dedicated system prompt and reduced length penalties during RL training .

This parameter allows developers to trade response time for reasoning depth, making the model adaptable for use cases from simple classification to advanced code generation and multi-step planning .

API Pricing: The Cheapest Frontier-Class API

DeepSeek V4 Flash is priced at $0.14 per million input tokens (cache miss) and $0.28 per million output tokens . For cached inputs (repeated system prompts or common prefixes), the price drops to $0.0028 per million tokens — a 98% reduction .

By comparison, V4 Flash output is 54x cheaper than Sonnet 4.6 ($15/M) and 107x cheaper than GPT-5.5 ($30/M) . Multiple sources describe it as "the cheapest frontier-class API available" .

Agentic and Coding Performance

The July 31 production release (V4-Flash-0731) delivered major agentic and coding gains through re-post-training rather than architectural changes . The official update log reports:

• Terminal Bench 2.1: 82.7
• NL2Repo: 54.2
• 76.7 on other agentic benchmarks

The release note states the model now "performs at comparable levels to its larger sibling V4-Pro on agentic and coding benchmarks" . This effectively flattens the traditional Pro/Flash performance hierarchy for agentic workloads .

DeepSeek Harness: The Agent Framework

DeepSeek is also pushing aggressively into agentic AI. On August 1, 2026, the company began recruiting open-source developers for the closed beta of DeepSeek Harness . The framework is designed to transform LLMs into AI agents that can autonomously execute tasks, manage context, and call tools .

The name first appeared in the July 31 V4-Flash-0731 changelog with the note "to be released soon" . The Harness engineering team was established in March 2026, when Cui Tianyi joined DeepSeek to build the team from scratch .

To participate in the beta, developers must have Agent Harness-related project experience and submit their GitHub ID along with representative work .

Strategic Context: Cheap, Capable, Open

DeepSeek's approach under CEO Liang Wenfeng centers on building cheap, capable models as a path to AGI . The V4 Flash pricing — $0.14/$0.28 per million tokens — and the MIT open-weight strategy are the clearest operational evidence of this philosophy. As one source notes, "V4-Flash at $0.28/M output is roughly 107× cheaper than GPT-5.5" .

DeepSeek faces domestic competition from Alibaba (Qwen series), ByteDance (Doubao), Moonshot AI, MiniMax, and Z.AI, but the V4 family is the only open-weight model lineup in its tier available under a permissive MIT license . Multiple sources also reference DeepSeek's reported IPO preparations, though specific dates or underwriters have not been confirmed .

What the Sources Don't Confirm

While the gathered sources paint a comprehensive picture, some claims could not be independently verified:

• Exact SWE-bench scores for V4-Flash-0731 were not found in the reviewed articles .
• Direct quotes from Liang Wenfeng on AGI strategy were not captured within the search budget.
• Specific IPO timeline or underwriters remain unconfirmed .
• Cui Tianyi's specific prior roles at TSY Capital and Jane Street are referenced in a single blog post but could not be independently corroborated from other sources .

These details may be available in broader industry reporting and could be confirmed with additional targeted searches.