DeepSeek V4 Flash 0731

DeepSeek V4 Flash 0731 is an efficiency-optimized Mixture-of-Experts (MoE) language model from DeepSeek, released on July 31, 2026 as the official version of the V4 Flash line, superseding the earlier April 2026 preview. As the lean tier of the V4 series, it retains 284 billion total parameters with 13 billion parameters activated per token, and continues to support a one-million-token context window while maintaining lower inference costs and memory overhead compared to the larger Pro variant.

Unlike the jump from V3 to V4, this update is a re-post-training of the same architecture and size, not a new design — DeepSeek's model card is explicit that the gains come from post-training changes rather than any architectural revision. Despite this, DeepSeek reports the 0731 release now outperforms both the original Flash preview and the V4 Pro (Preview) model across a range of agentic and coding benchmarks, and it is reported to sit close to frontier proprietary models on several agentic evaluations.

Architecture and Innovation

The model retains its Hybrid Attention Architecture, combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), which reduces KV cache requirements substantially compared to previous generations like DeepSeek-V3.2, keeping long-context processing accessible on standard hardware.

It also continues to use Manifold-Constrained Hyper-Connections (mHC) in place of conventional residual connections (expansion factor 4, with 20 Sinkhorn-Knopp iterations) to stabilize signal propagation across layers, and was pre-trained on 32+ trillion tokens using the Muon optimizer.

New in the 0731 checkpoint: the released weights ship with a DSpark speculative decoding module attached, matching the structure of the separate DeepSeek-V4-Flash-DSpark release. This pushes the Hugging Face repo's total parameter count to roughly 304B (284B base + draft module).

Reasoning and Max Effort

DeepSeek V4 Flash 0731 natively supports Thinking Mode, configurable via a reasoning_effort parameter with low, high, and max settings. "Think High" reasoning is on by default. Internal reasoning tokens remain accessible via API parameters such as reasoning_details.

DeepSeek recommends allowing up to 384K output tokens when using high or max reasoning effort, and suggests a minimum 384K context window for "Think Max" configurations. With enhanced post-training, the Flash model's reasoning performance in complex logic, mathematics, and programming tasks is now reported to exceed that of the V4 Pro (Preview) on several benchmarks, despite its far smaller activated parameter count.

Recommended sampling parameters:

• Agentic/local workloads: temperature = 1.0, top_p = 0.95
• Other workloads: temperature = 1.0, top_p = 1.0

Benchmarks (DeepSeek-reported, 0731 model card)

• Terminal-Bench 2.1: 82.7%
• DeepSWE: 54.4%
• NL2Repo: 54.2%
• DSBench-FullStack (internal): 68.7
• DSBench-Hard (internal): 59.6
• Artificial Analysis Intelligence Index: 50

Code Agent benchmarks were evaluated using the minimal mode of the (as-yet unreleased) DeepSeek Harness, at max reasoning effort, temperature = 1.0, top_p = 0.95. Because these agent scores are harness-sensitive, independent reproductions may vary.

Capabilities and Deployment

Optimized for coding assistants and complex autonomous agents, the model remains highly proficient in cross-file repository analysis and multi-step tool usage. It is available as open-source weights under the MIT license, supporting quantization formats including FP8 and FP4, and can run on a single-node setup for many long-context applications.

On the API side, deepseek-v4-flash now natively supports the Responses API format and has been adapted for Codex compatibility. Pricing is approximately $0.14 per million input tokens and $0.28 per million output tokens (figures vary slightly by host/provider due to caching and routing).