DeepSeek V4 Flash

DeepSeek's official API documentation announces that V4-Flash and V4-Flash-Vision-Exp are retired, and that the legacy "deepseek-v4-flash" and "deepseek-v4-flash-vision-exp" model names temporarily route to the newly released V4.1-Flash for backward compatibility. This is the only candidate that explicitly addresses th The same official post details that V4.1-Flash (the live target behind the retired deepseek-v4-flash alias) is a 552B-parameter MoE model with a new Causal Encoder–Decoder architecture, 8B active parameters for input and 16B for output, and a significantly smaller KV cache using 1/4 the HBM and 1/8 the SSD storage comp

DeepSeek announced DeepSeek-V4.1-Flash on September 9, 2026 as the new smallest model in its architecture family, but the post explicitly retires V4-Flash and V4-Flash-Vision-Exp, routing deepseek-v4-flash and deepseek-v4-flash-vision-exp to V4.1-Flash for compatibility. As of the announcement, V4-Flash is no longer se DeepSeek-V4.1-Flash is a 552-billion-parameter MoE with a new Causal Encoder–Decoder architecture using just 8 billion active parameters for input and 16 billion for output, native multimodal support, and a KV cache that needs roughly one-quarter the HBM and one-eighth the SSD storage of the previous generation. V4.1-F

The InferenceX model page for DeepSeek V4 provides technically substantive coverage of the V4 series, explicitly naming DeepSeek-V4-Flash as a 284B-total / 13B-active parameter MoE model co-released with the flagship DeepSeek-V4-Pro (1.6T total / 49B active) on April 23, 2026. The page cites DeepSeek's own API news rel While the page's architectural deep-dive focuses on V4-Pro (hybrid CSA + HCA attention, MLA-style shared-KV MQA, hash-routed MoE, FP4 experts with FP8 mixed precision, Muon optimizer, YaRN RoPE for 1M context), the explicit naming of V4-Flash with its parameter counts and shared release context provides direct model-le

According to a third-party API guide dated August 26, 2026, DeepSeek's hosted API exposes `deepseek-v4-flash` as a current production route, mapping via the official quick start to the snapshot build DeepSeek-V4-Flash-0731. The guide recommends V4 Flash as the default for text input and routine agent workloads, with V4 The guide explicitly distinguishes the DeepSeek-hosted `deepseek-v4-flash` route from the DeepSeek app, Hugging Face repositories, and similarly named gateway resales, framing API access as a separate contract that does not transfer to self-hosted quantizations. It also notes that the sibling `deepseek-v4-flash-vision-

Better Stack's technical guide confirms that DeepSeek-V4-Flash-0731 was published on July 31, 2026, moving the official V4-Flash API into public beta the same day. The model uses a sparse 284B-parameter MoE architecture with 13B parameters active per forward pass, a 1-million-token context window, a maximum output of 3 The guide documents launch pricing of $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens, and positions V4-Flash as occupying a useful point on the cost-intelligence curve rather than as a frontier-capability model. It includes a practical developer w

An independent model release tracker lists 22 DeepSeek releases between November 2023 and August 13, 2026, including the exact V4-Flash lineage relevant to the `deepseek-v4-flash` subject: DeepSeek-V4-Flash shipped on April 24, 2026, and was updated to DeepSeek-V4-Flash-0731 on July 31, 2026. Both entries are recorded The tracker surfaces benchmark context that contextualizes V4 Flash within the V4 family, even though Flash-specific scores are not broken out: V4-Pro-0813 records 80.6% on SWE-Bench Verified, 31.8% on AutomationBench, 87.9% on Terminal-Bench 2.1, and 60% on Humanity's Last Exam, with the earlier V4-Pro posting 93.5% o

DeepSeek published V4-Flash-0731 on July 31, 2026, a retrained checkpoint of the 284-billion-parameter Mixture-of-Experts model first previewed in April, with the architecture unchanged at 13 billion active parameters per token. According to the article, the improvements come entirely from an expanded reinforcement-lea The article reports that V4-Flash-0731 beats DeepSeek's own V4-Pro-Preview on nine published agent and coding benchmarks, with pricing positioned at $0.14 per million input tokens and $0.28 per million output tokens. The author flags reproducibility asterisks on the headline benchmark numbers and frames the broader sig

DeepSeek released DeepSeek-V4-Flash-0731 on July 31, 2026, as the official V4-Flash build superseding the April 2026 preview. It is a 284-billion-parameter Mixture-of-Experts model with 13 billion activated parameters per token and a 1 million-token context window. The checkpoint on Hugging Face reports 304 billion par DeepSeek reports the 0731 build beats the larger V4-Pro preview on every published agentic benchmark, with a Terminal-Bench 2.1 score of 82.7 (up from 61.8 on the April preview). The deepseek-v4-flash API endpoint now natively supports the Responses API format and has been adapted for Codex, while the V4-Pro API and th

DeepSeek V4-Flash entered official public beta on July 31, 2026 as build V4-Flash-0731, retaining the same 284B-parameter architecture, context window, and price as the April preview. Across all nine agent and coding benchmarks DeepSeek published, the build outscores its own V4-Pro preview, including a jump from 7.3 to The rollout covers only the hosted API and not the public Hugging Face repository, which still hosts April's weaker preview build, leaving self-hosters on the older checkpoint. V4-Pro's official release is expected in August 2026, and DeepSeek's API changelog notes the calling method is unchanged, with deepseek-v4-flas

DeepSeek released the official DeepSeek-V4-Flash API into public beta on July 31, 2026, under the model identifier "deepseek-v4-flash," with the release labeled V4-Flash-0731. According to the article, V4-Flash-0731 scored 82.7 on Terminal Bench 2.1, reportedly outperforming DeepSeek's own V4-Pro-Preview (72.1) by 14.7 Pricing is reported at $0.14 per million input tokens, which the article frames as the most cost-effective frontier agent model on the market, with full backward-compatible API access via the "deepseek-v4-flash" model name. The post discusses the implications for AI agent development and notes support for the Responses

DeepSeek-V4-Flash-0731 was released on July 31, 2026, 98 days after the original V4-Flash, and is tracked against benchmarks including SWE-Bench Verified, Terminal-Bench 2.1, BrowseComp, and Toolathlon-Verified. The release tracker documents verified API pricing pulled from api-docs.deepseek.com, distinguishing off-pea The page records peak-hour pricing (01:00–04:00 and 06:00–10:00 UTC) at $0.44 input / $0.014 cached / $1.32 output per million tokens, with off-peak rates at $0.22 / $0.007 / $0.66. Thinking and non-thinking modes use the same published price, and rates were verified against DeepSeek's official API documentation on Aug

DeepSeek V4 Flash 0731 scored 50 on the Artificial Analysis Intelligence Index, a 10-point jump over the original DeepSeek V4 Flash released in April 2026, and 6 points above DeepSeek V4 Pro. It lands within 1 point of GLM-5.2 (max) and Gemini 3.6 Flash (50), placing it near the frontier of cost-efficient open-weight m Agentic gains are the headline: GDPval-AA v2 Elo rose to 1559 from 1189 (second-highest open-weights score, behind Kimi K3 and ahead of GLM-5.2), Terminal-Bench 2.1 climbed 17 points to 79%, and τ³-Bench Banking improved 8 points to 31%. Omniscience gains came from reduced hallucination: the AA-Omniscience Index improv