DeepSeek Upgrades V4-Flash 0731 with Major Agentic and Coding Improvements

DeepSeek has rolled out an update for the DeepSeek-V4-Flash model, designated as checkpoint 0731. This release ships with the DSpark speculative decoding module attached, matching the structural setup of DeepSeek-V4-Flash-DSpark. According to Hugging Face reports, the repository contains 304B parameters in total, which accounts for the draft module stacked on top of the 284B base model.

On the API front, deepseek-v4-flash now natively supports the Responses API format and has been adapted for Codex, while the V4-Pro API alongside app and web models remain unupdated in this cycle. Deployment via API is remarkably accessible: DeepSeek’s pricing page lists deepseek-v4-flash at $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens, with a concurrency limit set at 2,500. This pricing sits at roughly a third of the deepseek-v4-pro output pricing which stands at $0.87, allowing seed-stage startups, indie developers, and internal platform teams to execute agent loops without requiring a heavy GPU budget.

For self-hosting scenarios, the technical requirements are significantly higher. Although the model weights are MIT-licensed and ungated, every expert must stay resident in memory even though only 13B parameters activate per token. DeepSeek’s vLLM reference example serves the model on a single 4×GB300 node, whereas Unsloth’s dynamic GGUFs require approximately 110 GB of combined RAM and VRAM for the lossless 8-bit build at 162 GB and the 3-bit build at 103 GB. Consequently, self-hosting is best suited for mid-size and large enterprises equipped with dedicated serving clusters or high-spec workstations running aggressive quantization.

Stock photo for illustration only, not from the actual event

According to the DeepSeek-V4 technical report, V4-Flash is a 284B-parameter Mixture-of-Experts (MoE) architecture featuring 13B activated parameters per token and a 1M-token context window capacity. Each MoE layer contains 1 shared expert and 256 routed experts with an intermediate dimension of 2048, with 6 routed experts firing per token. The initial three MoE layers utilize hash routing, and the multi-token prediction depth is set to 1.

The attention mechanism is hybrid, combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Manifold-Constrained Hyper-Connections (mHC) replace traditional residual connections, utilizing an expansion factor of 4 and 20 Sinkhorn-Knopp iterations. Pre-training utilized over 32T tokens alongside the Muon optimizer. Note that the paper's headline efficiency figures—achieving 27% single-token inference FLOPs and 10% KV cache compared to DeepSeek-V3.2 at a 1M context—apply specifically to V4-Pro rather than Flash, with all figures sourced directly from DeepSeek's 0731 model card.

Enabling DSpark requires just a single vLLM flag:
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
The DSpark research paper indicates a 60–85% increase in per-user generation speed on V4-Flash compared to the MTP-1 baseline under matched aggregate throughput conditions.

Regarding integration details, there is no Jinja chat template included. Instead, DeepSeek provides an encoding/ folder containing encode_messages and parse_message_from_completion_text functions. The reasoning_effort parameter accepts low, high, or max values. DeepSeek recommends setting temperature = 1.0 and top_p = 0.95 for agentic use cases and 1.0 otherwise, supporting up to 384K output tokens when configured to high and max settings.

Source: MarkTechPost