DeepSeek-V4-Flash-0731

DeepSeek-V4-Flash-0731 is DeepSeek's July 31, 2026 GA release of its efficiency-tier MoE model, holding a 1,048,576-token context window and Terminal-Bench 2.1 score of 82.7. It suits cost-sensitive coding agents and long-context pipelines more than latency-critical chat or confirmed vision workloads.

DeepSeek-V4-Flash-0731, released July 31, 2026, is a 284-billion-parameter Mixture-of-Experts model with 13 billion active parameters, an 88.1 GPQA Diamond score, and a 1,048,576-token context window. It re-post-trains the April preview for agentic coding, lifting DeepSWE 7.5x with no architecture change.

Provider: DeepSeek · Family: DeepSeek V4

More about DeepSeek on HokAI

Context window: 1,048,576 tokens · Max output: 65,536

Input modalities: text, tool-calls · Output: text, tool-calls

About DeepSeek-V4-Flash-0731

DeepSeek-V4-Flash-0731 is the official, generally-available release of DeepSeek's efficiency-tier Flash model, announced by DeepSeek on July 31, 2026. It supersedes the DeepSeek-V4-Flash-Preview build that first shipped April 24, 2026, keeping the same architecture and parameter count while swapping in a substantially improved post-training pipeline focused on coding, agentic tool use, and reasoning. It is a sparse Mixture-of-Experts model with 284 billion total parameters and 13 billion activated per token, using 1 shared expert plus 256 routed experts (6 fired per token, intermediate dimension 2048). Attention uses a hybrid of Compressed Sparse Attention and Heavily Compressed Attention for long-context efficiency, and residual connections are replaced with Manifold-Constrained Hyper-Connections (expansion factor 4, 20 Sinkhorn-Knopp iterations). The post-training pass produced large benchmark jumps without touching the underlying weights' size or shape. DeepSWE rose from 7.3 in the preview to 54.4, a 7.5x gain, and Terminal-Bench 2.1 climbed from 61.8 to 82.7, a 20.9 point jump. On broader academic evals, DeepSeek's own reporting and third-party model-card discussion place GPQA Diamond at 88.1 and MMLU-Pro at 86.4, with HumanEval around 90 and MMLU around 89. For comparison, DeepSeek's own larger V4-Pro-Max sibling scores 90.1 on GPQA Diamond and 87.5 on MMLU-Pro, while Gemini 3.1 Pro reports 94.3 GPQA Diamond and 91.0 MMLU-Pro. These vendor-reported numbers run on an unreleased internal harness and have not been independently reproduced at time of writing, so treat them as directionally accurate rather than final. Context window is 1,048,576 tokens (roughly 1M), with a maximum output commonly cited at 65,536 tokens, though some third-party docs list higher output ceilings depending on provider configuration. DeepSeek did not disclose a formal system card alongside the 0731 release, and independent latency figures are limited: community benchmarking on OpenRouter puts throughput around 93 tokens per second, workable for agent loops but not the fastest option for latency-sensitive chat. Model inputs and capabilities center on text: tool use, function calling, and step-by-step reasoning are all supported and were the explicit target of the 0731 re-training. Vision input is reported inconsistently across sources as of this writing (some list image support, others describe it as text-only), so it should be verified per-provider before depending on it in production. The 0731 build adds native Anthropic-style Responses API support and was specifically adapted for coding agents such as Codex-style harnesses, which is consistent with the DeepSWE and Terminal-Bench jumps. Pricing held steady through the GA release at $0.14 per 1M input tokens and $0.28 per 1M output tokens, among the cheapest frontier-adjacent models on the market and roughly 20-40x cheaper per token than Western frontier flagships. DeepSeek's own API is the only provider currently serving the true 0731 checkpoint; OpenRouter, Fireworks, Vercel AI Gateway and other routers list DeepSeek V4 Flash but some still point at the older 0424 preview weights rather than 0731, so provider selection matters for anyone chasing the newest benchmark gains. On the open-weights side, DeepSeek published MIT-licensed weights for the V4-Flash line on Hugging Face, pre-quantized in mixed FP4 (MoE experts) plus FP8 (attention, norm, router) precision. Full-precision inference of the native weights needs roughly 170-175GB of VRAM, fitting on 2x H200 or 2x RTX Pro 6000 Blackwell cards. Community-produced GGUF and MLX quantizations bring that down substantially: an 8-bit lossless build runs around 162GB, and a 3-bit build drops to roughly 103GB, needing about 110GB of combined RAM plus VRAM for local inference. Some reporting notes the downloadable MIT weights track the April preview checkpoint rather than the exact 0731 post-training run, so self-hosted deployments may not match the hosted API's benchmark scores exactly. The model is best suited to cost-sensitive coding agents, high-volume tool-calling pipelines, and long-document workflows that can tolerate a mid-pack (not fastest) throughput in exchange for a 1M-token window and sub-$0.15 input pricing. It is a weaker fit for latency-critical consumer chat, workloads that need a confirmed native vision pipeline, or teams that require a disclosed safety system card and red-team documentation before deployment, since DeepSeek has not published one for this specific build. Teams weighing it against GPT-5-class or Claude Opus-class models should expect comparable or slightly lower academic benchmark scores at a small fraction of the per-token cost, with the tradeoff being thinner published safety documentation and inconsistent third-party vision support. DeepSeek's release cadence for the V4-Flash line has been rapid: preview in April 2026, GA re-post-train in July 2026, roughly a 3-month cycle. Given that pattern, further post-training refreshes without an architecture change are plausible before a true V5 generation, so teams pinning to a specific checkpoint should track the exact build suffix (0424 vs 0731) rather than assuming "DeepSeek V4 Flash" always means the newest weights.

Pricing

$0.14 per 1M input tokens, $0.28 per 1M output tokens on DeepSeek's own API, held flat through the July 31, 2026 GA release. Some third-party routers (OpenRouter, ZenMux) offer free-tier variants pointing at older preview weights.

Key Features

• 1M-Token Context Window: 1,048,576-token context window supports full-document and multi-file codebase ingestion in a single request.
• Agentic Re-Post-Train: The 0731 GA release re-trains the same architecture on coding and tool-use data, lifting DeepSWE 7.5x and Terminal-Bench 2.1 by 20.9 points versus the April preview.
• MIT-Licensed Open Weights: Weights published on Hugging Face under MIT license, with native FP4/FP8 quantization and community GGUF/MLX builds for local inference.
• Sparse MoE Efficiency: 284B total parameters with only 13B active per token, keeping inference cost low relative to dense models of similar capability.
• Sub-$0.15 Input Pricing: $0.14 per 1M input tokens and $0.28 per 1M output tokens on DeepSeek's own API, among the cheapest long-context models available.

Pros

• DeepSWE score jumped 7.5x (7.3 to 54.4) from post-training alone, with no architecture change.
• 1,048,576-token context window at $0.14 per 1M input tokens undercuts most frontier-adjacent competitors on cost.
• MIT-licensed open weights give self-hosting teams a permissive-license path with native FP4/FP8 quantization.

Cons

• No published system card or disclosed red-team partners for this specific build.
• Vision input support is reported inconsistently across providers; not safe to assume without per-provider testing.
• Only DeepSeek's own API confirms the exact 0731 checkpoint; several third-party routers still serve the older 0424 preview under the same model name.

Benchmarks

• mmlu: 89
• deepswe: 54.4
• mmlu pro: 86.4
• humaneval: 90
• gpqa diamond: 88.1
• terminal bench 2 1: 82.7
• artificial analysis price blended per m: 0.196
• artificial analysis speed tokens per sec: 93

Visit DeepSeek-V4-Flash-0731 Official Page