DeepSeek-V4-Flash

September 10, 2026 update: V4 Flash 0731

This page preserves DeepSeek V4 Flash as a distinct model with its existing lifecycle status. No retirement, deletion, or model-page redirect is applied. The V4 Flash 0731 column in the supplied September 10 release table updates its comparison results; it must not be confused with V4.1 Flash.

Updated max-effort results

GPQA Diamond 89.9; HLE pure-text subset 37.8; Codeforces rating 3289; MathArena Apex 58.6; Terminal-Bench 2.1/3.0/4.0: 82.7/7.6/7.0; DeepSWE v1.1 54.4; NL2Repo-Bench 54.2; CyberGym 76.7; SEC-Bench Pro 30.9; ExploitGym 1.8 (budget unspecified); HLE with tools 51.5; Automation-Bench 37.7; Agents’ Last Exam 25.2.

The supplied clarification identifies max effort, with tool use tracked separately and internet access not inferred. Pure-text HLE is kept separate from full HLE, and budget-unspecified ExploitGym is separate from 2h/6h/unlimited protocols. Dashes for ProgramBench and visual evaluations mean no result supplied, not zero. The original 0731 Toolathlon Verified score of 70.3 and other historical modes and third-party results are retained. The new Automation-Bench value 37.7 replaces 25.1 in the max-with-tools entry; the old release value remains documented in the local update record.

Architecture and open-weight identity

Existing specifications remain: 284B total and 13B active parameters, MoE, text-only input and output, 1M context and 384K maximum output. The V4 architecture combines Compressed Sparse Attention, Heavily Compressed Attention, mHC connections and the Muon optimizer. The original V4 Flash weights and their existing MIT license remain distinct from the new API route. The 0731 release updated API post-training, not the previously published weights.

Compatibility routing and historical prices

According to the supplied September 10 materials, deepseek-v4-flash temporarily routes to V4.1 Flash; the recommended API name is deepseek-flash. Calls through the compatibility name execute the new model and use its peak/off-peak prices. This does not alter the original model’s lifecycle, weights, or historical benchmarks.

The price rows retained for the original model are historical: USD 0.0028 cache-hit input, 0.14 cache-miss input, and 0.28 output per million tokens. They are not current quotes for requests routed through the compatibility alias. V4.1 Flash separately lists off-peak CNY 0.02/1/4 and peak CNY 0.04/2/8 for those three categories. Peak hours are Monday–Friday 09:00–12:00 and 14:00–18:00 Asia/Shanghai.

Historical reasoning and API support

Original off/high/max thinking modes and the existing default are preserved. Original API support includes OpenAI/Anthropic compatibility, Responses API, JSON Output, Tool Calls, prefix completion, and non-thinking FIM. V4 Flash remains a text model; routing its old API name to a vision model does not give its own weights image support.

Source and evaluation notes

Updated scores and routing information come from the supplied DeepSeek September 10, 2026 release materials without independent source verification in this update. Original 0731 public Code Agent notes specified DeepSeek Harness minimal mode, max effort, temperature=1.0 and top_p=0.95; this does not establish identical settings for every historical benchmark. Reference: DeepSeek change log.