deepseek-ai/DeepSeek-V4-Flash (official card, 'Comparison across Modes' instruct table)

Fields

id

sources/deepseek-v4-flash-hf-card

URL

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash

Kind

repo

Accessed

2026-10-19

notes

Read https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash for the benchmark figures on models/deepseek-v4-flash. As recorded by the reader: 284B total / 13B activated confirmed on the card. Reported values are the V4-Flash 'Max' reasoning-effort instruct column (MMLU-Pro EM 86.2; GPQA Diamond Pass@1 88.1; SWE-bench Verified Resolved 79.0). Same table other modes: High 86.4/87.4/78.6 and Non-Think 83.0/71.2/73.7. Base-model row values were not used. Card's livecodebench row is unbounded 'LiveCodeBench' (91.6) and its agentic rows are Terminal Bench 2.0 / SWE Pro, so those were excluded; the separate DeepSeek-V4.1-Flash card reports different (higher-effort) DS-V4-Flash numbers (GPQA 89.9, Terminal-Bench 2.1 82.7) which are not used here to avoid conflicting values for this slug.

Notes

Read https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash for the benchmark figures on models/deepseek-v4-flash. As recorded by the reader: 284B total / 13B activated confirmed on the card. Reported values are the V4-Flash 'Max' reasoning-effort instruct column (MMLU-Pro EM 86.2; GPQA Diamond Pass@1 88.1; SWE-bench Verified Resolved 79.0). Same table other modes: High 86.4/87.4/78.6 and Non-Think 83.0/71.2/73.7. Base-model row values were not used. Card's livecodebench row is unbounded 'LiveCodeBench' (91.6) and its agentic rows are Terminal Bench 2.0 / SWE Pro, so those were excluded; the separate DeepSeek-V4.1-Flash card reports different (higher-effort) DS-V4-Flash numbers (GPQA 89.9, Terminal-Bench 2.1 82.7) which are not used here to avoid conflicting values for this slug.

Cross-references

This record references no other records.