---
title: "DeepSeek-V4-Flash Hits 79% on SWE-bench at $0.14 per Million Tokens"
subtitle: "DeepSeek V4 Flash brings 1M-token context and frontier-adjacent reasoning at $0.14/M input tokens, under MIT license with open weights"
url: "https://alphasignal.ai/news/deepseek-v4-flash-hits-79-on-swe-bench-at-0-14-per-million-tokens"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "DeepSeek"
company_url: "https://alphasignal.ai/company/deepseek_ai"
published: "2026-04-22T06:04:20.000Z"
updated: "2026-07-04T02:51:22.462Z"
type: "Model"
topics:
  - "Llms"
  - "Open Source"
subtopics:
  - "Long Context"
  - "Mixture Of Experts"
  - "Small Models"
upvotes: 4286
source_url: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
comments: 0
truncated: true
---

# DeepSeek-V4-Flash Hits 79% on SWE-bench at $0.14 per Million Tokens

> DeepSeek V4 Flash brings 1M-token context and frontier-adjacent reasoning at $0.14/M input tokens, under MIT license with open weights

[DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) is the smaller, faster, and dramatically cheaper half of DeepSeek's new V4 model family. It is not a distilled or stripped-down version of the flagship Pro model. It was trained from scratch on the same dataset, with the same architectural innovations, and it punches well above its weight class on benchmarks that matter for real production workloads.

The V4 family consists of two Mixture-of-Experts (MoE) models: DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated), both supporting a context length of one million tokens. The key number for Flash is that 13B figure. In a MoE architecture, only a fraction of the total parameters are actually used for any given token, which is what makes the model cheap to run despite its massive total size.

## A new architecture, not just more parameters

V4 Flash and V4 Pro represent a new architecture, not just an incremental update. V4 is not V3.2+. It is a new architecture with a fundamentally different approach to long context. The three core innovations are worth understanding because they directly affect what you can do with the model.

* **Hybrid Attention (CSA + HCA):**DeepSeek designed a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. CSA compresses the sequence 4x and uses a top-k indexer to find relevant tokens; HCA compresses 128x into a dense stream with a sliding window for recency. Together, they let the model process very long contexts without the usual memory explosion.
* **Manifold-Constrained Hyper-Connections (mHC):**mHC strengthens conventional residual connections, enhancing stability of signal propagation across layers while preserving model expressivity. Residual connections are the skip-connections that let signals bypass layers; mHC constrains these to a geometric manifold, making training more stable.
* **Muon Optimizer:**The Muon optimizer is employed for faster convergence and greater training stability. This replaces the standard AdamW optimizer used in most large model training runs.

At 1M tokens, V4-Pro uses only 27% of the per-token FLOPs and 10% of the KV cache of V3.2; Flash drops those to 10% and 7%. That efficiency gain is what makes a 1M-token context window economically viable to serve at scale, rather than just technically possible.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [DeepSeek-V4-Flash Model Card - Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash)
- [DeepSeek V4 Preview Release - Official API Docs](https://api-docs.deepseek.com/news/news260424)
- [DeepSeek V4 in the wild, and how to run it on Runpod](https://www.runpod.io/blog/deepseek-v4-in-the-wild-and-how-to-run-it-on-runpod)
- [DeepSeek V4 Flash - API Pricing & Benchmarks - OpenRouter](https://openrouter.ai/deepseek/deepseek-v4-flash)
- [DeepSeek V4 Guide: Pro & Flash (May 2026)](https://codersera.com/blog/deepseek-v4-complete-guide-2026/)