DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is positioned as the smallest entry in a new model architecture series from DeepSeek, designed to bring capable language and agent performance into a lighter footprint. According to DeepSeek's official release announcement, the model features comprehensive improvements in text and Agent performance alongside native multimodal visual understanding, allowing it to interpret images as part of its input stream rather than relying on a bolted-on vision adapter. This combination of compact design and built-in visual reasoning makes it well suited for agent-driven workflows where both textual context and visual information need to be processed in a single pass.

On the API side, DeepSeek V4.1 Flash is offered under the model name deepseek-flash, replacing the earlier V4 Flash and V4 Flash Vision Exp variants. DeepSeek introduced a Peak-Valley pricing structure alongside the release, lowering prices through architecture innovation and setting off-peak rates at half of the peak-period rate. The model is intended as a fast, cost-efficient option for developers building assistants, tool-using agents, and multimodal applications where quick response times and lower inference costs are more important than the largest possible context or the deepest reasoning capacity.

Synthetichf:deepseek-ai/DeepSeek-V4.1-Flashdeepseek-flash

Open provider page Provider docs

Quick Info

Provider: Synthetic
Model key: hf:deepseek-ai/DeepSeek-V4.1-Flash
Release date: Sep 10, 2026
Last updated: Sep 10, 2026
Knowledge cutoff: 2025-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.60
Output token cost: $1.20

Limits

Output tokens: 65,536 tokens
Context window: 524,288 tokens

Transparent token rates

Compare DeepSeek V4.1 Flash pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about DeepSeek V4.1 Flash

Synthetic

CoverageBenchmark

Eesel AI's analysis confirms DeepSeek shipped V4.1 Flash on 10 September 2026 and quotes DeepSeek calling it "the smallest model in our new architecture family, with native visual understanding," framing it as a new design rather than a minor tweak. It documents the model-family consolidation: the old deepseek-v4-flash The article discusses the architectural and pricing implications of V4.1 Flash for production users, including how KV-cache compression and MoE efficiency changes affect deployment cost, and surfaces community reactions on Hacker News about pinning models in production when providers swap implementations on paying cust

Synthetic

CoverageBenchmark

Flowtivity's benchmark analysis reports that DeepSeek V4.1 Flash, released 10 September 2026, beats OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus-5.0 on four of five hard agentic benchmarks: DeepSWE v1.1 (74.2 vs 73.0), AutomationBench (54.8 vs 45.8), Agents' Last Exam (31.8 vs 26.7), and CyberGym (88.1 vs 84.5). Th The piece also explains the KV-cache economics behind V4.1 Flash's agentic wins, notes where the model still loses against frontier peers, and details the housekeeping around prior DeepSeek variants. DeepSeek is retiring the roughly four-times-more-expensive V4 Pro on 14 September 2026, with the company stating that "V

Synthetic

CoverageBenchmark

LLM Stats' DeepSeek-V4.1-Flash profile aggregates the model's benchmark performance, ranking it 13 on the LLM Stats Score composite. The page reports a Codeforces competitive-programming rating of 3471 (with reasoning effort=100, temperature=1.0, top-p=0.95), a GPQA Diamond Pass@1 of 0.91, and a Terminal-Bench 2.1 Pass The profile also positions DeepSeek-V4.1-Flash on cost efficiency, plotting its blended price per million tokens (around $0.24) against its LLM Stats Score (51.4) alongside neighboring models such as Gemma 4 E4B, GPT OSS 120B, and DeepSeek-V4-Flash-0731, giving developers a quick reference for how V4.1 Flash compares o

Synthetic

Coverage

The alphaXiv paper page introduces DeepSeek-V4.1-Flash as a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts up to one million tokens, designed for input-heavy long-horizon agent workloads where prefill and large KV caches dominate cost. Its Causal Encoder-Decoder (CED) archite The paper details the KV-cache compression innovations at the heart of V4.1-Flash: cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) combined with FP4 KV caching reduces the global KV cache footprint in HBM to 890 bytes per token, roughly one-quarter of DeepSeek-V4-Flash's footprint, while a SWA Bounde

Synthetic

Coverage

DeepSeek's official API documentation Change Log dated 2026-09-10 announces the release of DeepSeek-V4.1-Flash, described as the smallest model in a new architecture family with native multimodal visual understanding, designed for higher capability ceiling, faster inference, and throughput scaling. The model can now be The Change Log lists DeepSeek-V4.1-Flash benchmark scores directly from the official release: GPQA Diamond 90.9, HLE 36.8 (39.1 on the pure-text subset), Codeforces rating 3471, MathArena Apex 65.6, Terminal-Bench 2.1 90.6, Terminal-Bench 3.0 30.0, Terminal-Bench 4.0 31.2, DeepSWE v1.1 74.2, ProgramBench 20.3, NL2Repo-