Model

DeepSeek-V4-Flash-0731

OPEN

DeepSeek · DeepSeek-V4 family · released Jul 31, 2026

Retrained checkpoint of DeepSeek-V4-Flash published July 31, 2026, same 284B/13B architecture and API pricing as the April siblings. DeepSeek's own model card only reruns the launch-day agentic/coding table (Terminal-Bench 2.1, NL2Repo, CyberGym, DeepSWE) with an unreleased internal harness ('DeepSeek Harness'); general reasoning/knowledge benchmarks were not rerun and still carry April Preview values, so this row is coding/agentic-only evidence, not a full refresh.

ReasoningCodingVisionFunction callingTool useAgentic

2777.6

Elo · unrated

Parameters

284B

Active params

13B (MoE)

Context

1M tokens

Architecture

Mixture-of-Experts (284B total / 13B active), hybrid Compressed Sparse Attention + Heavily Compressed Attention — retrained checkpoint of DeepSeek-V4-Flash on the same architecture

License

MIT

Languages

—

API price (in/out)

$0.14 / $0.28

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

CyberGymCoding76.7%#6

best: GPT-5.6 Sol · 84.5%

DeepSWECoding54.4%#5

best: GPT-5.6 Sol · 72.7%

Terminal-Bench 2.0Coding82.7%#9

best: GPT-5.6 Sol · 88.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA193.1 GB2× H200 141GB

LoRA604.9 GB4× B200 192GB

Full fine-tune4558.2 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $6.72 (2× H200 141GB)

Plan a fine-tune →

DeepSeek-V4 family

Elo progression across releases

DeepSeek-V4-Flash (Non-Think)Apr 20261632.3 DeepSeek-V4-Flash (Think High)Apr 20262234.2 DeepSeek-V4-Flash (Think Max)Apr 20262319.9 DeepSeek-V4-Pro (Non-Think)Apr 20261676.6 DeepSeek-V4-Pro (Think High)Apr 20262386.5 DeepSeek-V4-Pro (Think Max)Apr 20262511.5 DeepSeek-V4-Flash-0731Jul 20262777.6

Resources

HFdeepseek-ai/DeepSeek-V4-Flash-0731 DOCShttps://api-docs.deepseek.com/quick_start/pricing/APIAvailable via DeepSeek API

API price $0.14/$0.28 · each benchmark row carries its own source badge (see methodology)