Model
Model explorer

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. DeepSeek retired V4 Flash on Sep 10, 2026; the deepseek-v4-flash name now routes to V4.1-Flash.

ReasoningCodingVisionFunction callingTool useAgentic
2528.2
Elo · rank #40
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
ARC-AGI-1Reasoning89.0%#23
best: Claude Fable 5 · 98.5%
ARC-AGI-2Reasoning61.4%#25
best: GPT-6 Astra · 95.0%
CyberGymCoding76.7%#13
best: MiMo-V2.6-Flash · 95.1%
DeepSWECoding54.4%#29
best: Muse Wonder Full · 79.0%
best: Muse Spark 1.2 · 25.4%
NL2RepoCoding54.2%#6
best: DeepSeek-V4.1-Flash-1017 · 68.2%
SWE-bench VerifiedCoding88.8%#10
best: Claude Opus 5 · 97.0%
Terminal-Bench 2.0Coding82.7%#18
best: DeepSeek-V4.1-Flash · 90.6%
Terminal-Bench 4.0Agents18.7%#31
best: Claude Sonnet 5.5 · 70.6%
best: GPT-6 Astra · 62.9%
Vals Finance AgentAgents49.5%#36
best: Gemini 4 Argon · 65.4%
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)
API price $0.14/$0.28 · each benchmark row carries its own source badge (see methodology)