DeepSeek V4 Flash 0731

The Hugging Face model card for DeepSeek-V4-Flash-Vision-Exp introduces it as the first experimental multimodal model in the DeepSeek-V4 family, building on the V4-Flash architecture with added visual modules and continued training to unlock visual understanding. On text-agent benchmarks (Terminal Bench 2.1 83.9, NL2Re The repository layout section confirms the release ships the tokenizer, prompt encoding reference, and a minimal PyTorch inference implementation covering the vision encoder and aligner, DFlash attention, MoE, Hyper-Connections, and the DSpark forward path, deliberately split between encoding/ and inference/ directorie

GIGAZINE reported on August 3, 2026 that DeepSeek released DeepSeek-V4-Flash-0731 as an open model under a license permitting commercial use, with weights hosted on Hugging Face and ModelScope. The piece describes it as an updated version of DeepSeek-V4-Flash from April 2026, retaining the same MoE architecture with 28 Citing Artificial Analysis intelligence tests, the article places DeepSeek-V4-Flash-0731 at performance equivalent to Google's Gemini 3.6 Flash, ranking third among open models behind Kimi K3 and GLM-5.2. Coding performance is reported to surpass GLM-5.2, and agent performance is shown to significantly outperform Gemin

A Medium summary dated August 1, 2026 reports that DeepSeek officially announced DeepSeek-V4-Flash-0731 on July 31, 2026 as the production-ready version of its Flash model. The piece emphasizes that the release keeps the same 284 billion total parameters, 13 billion active parameters per token via Mixture-of-Experts, 1 According to the article, the gains come from post-training improvements rather than architectural changes, including improved reinforcement learning, better supervised fine-tuning, multi-teacher distillation, and stronger agent-focused training datasets. DeepSeek claims significantly better reasoning, coding, and AI a

On July 31, 2026, DeepSeek officially announced DeepSeek-V4-Flash-0731 as the production (General Availability) release of its Flash model, according to a third-party write-up by Mehul Gupta. The article states the model retains the same architecture as the earlier April Preview: 284 billion total parameters with 13 bi The piece frames DeepSeek-V4-Flash-0731 as evidence of an industry shift toward better post-training pipelines as a lever for capability gains, rather than simply enlarging model size. It is authored by an independent Medium commentator who paraphrases DeepSeek's announcement rather than linking to a primary release pa

OfficeChai reported on July 31, 2026 that DeepSeek pushed the official API for DeepSeek-V4-Flash-0731 into public beta, with the company making a substantial upgrade to the model's agentic capabilities and benchmark scores now ahead of its own V4-Pro-Preview. The API now natively supports the Responses format and has b The article highlights that V4-Flash-0731 keeps the same architecture and parameter count as the preview, with the update applying only to the V4-Flash API while V4-Pro and the consumer app/web versions remain on their existing builds. Benchmark jumps include Terminal Bench 2.1 to 82.7 (from 61.8 for the preview), NL2R

Artificial Analysis published a benchmark analysis on July 31, 2026 showing DeepSeek V4 Flash 0731 scoring 50 on the Artificial Analysis Intelligence Index, a 10-point jump over the earlier DeepSeek V4 Flash (40) and 6 points above DeepSeek V4 Pro. The model keeps identical architecture and pricing as the prior Flash, The analysis confirms DeepSeek V4 Flash 0731 retains a 1M token context window with 284B total parameters and 13B active at inference. Key gains are in agentic work: GDPval-AA v2 Elo climbs to 1559 (from 1189), Terminal-Bench 2.1 rises 17 points to 79%, and τ³-Bench Banking rises 8 points to 31%. AA-Omniscience Index i

DeepSeek shipped DeepSeek-V4-Flash-0731 on 31 July 2026 as the official GA release of the V4-Flash variant, superseding the April preview at the same initial price (peak/off-peak pricing was later introduced on 16 August 2026). The model is an open-source Mixture-of-Experts release with 284 billion total parameters and DeepSeek's own model card places V4-Flash-0731 at 82.7 on Terminal Bench 2.1, up from 61.8 for the preview and above the larger V4-Pro preview's 72.1, while Artificial Analysis moved the Flash from 40 to 50 on its Intelligence Index, placing it third in the open-weight field. Note: this item describes the 0731 checkpoi

A Hacker News discussion thread on the Artificial Analysis analysis of DeepSeek V4 Flash 0731 attracted 594 points and 312 comments, providing community signal around the July 31, 2026 release. The thread links to the open-weights repository at huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 and discusses the model's Commenters noted that DeepSeek's evaluation used the minimal mode of DeepSeek Harness (to be released) as the agent framework with max reasoning effort, fueling speculation that DeepSeek may release an optimized coding agent harness alongside the open weights. The discussion positions V4 Flash 0731 as a strong daily-dr

OpenRouter's page for deepseek/deepseek-v4-flash-0731 confirms the July 31, 2026 release of DeepSeek V4 Flash 0731 as a sparse mixture-of-experts model with 284B total parameters and 13B active, a 1M token context window, and a stated base price of $0.05 input / $0.16 output per 1M tokens with text in/out modalities. D As a router, OpenRouter enumerates many downstream inference hosts (OpenInference, Relace, DigitalOcean, DeepInfra, Together, Baseten, CoreWeave, Fireworks, Cloudflare, DeepSeek first-party, and others) with per-host latency, throughput, uptime, and pricing tiers for serving DeepSeek V4 Flash 0731. These figures reflec

The DeepSeek API change log dated 2026-08-21 documents the release of DeepSeek-V4-Flash-Vision-Exp, the first multimodal vision understanding model in the V4 family, accessible via model='deepseek-v4-flash-vision-exp'. Reported scores include Terminal Bench 2.1 at 83.9, NL2Repo 57.7, DeepSWE 59.3, DSBench-Hard 63.6, Au It also confirms that V4-Pro and V4-Flash now expose three thinking effort levels — low, high, and max — giving developers finer-grained control over deliberation cost. Vision-Exp is labeled experimental and is accessed through the same DeepSeek API platform, with usage details pointed to a dedicated Vision guide. This