DeepSeek drops V4-Flash: Agent muscle that beats its own Pro model

DeepSeek-V4-Flash-0731 is an open-source model from DeepSeek that officially replaces its earlier preview build, targeting high-efficiency autonomous agent and coding tasks.

Under the hood, the release adopts the DeepSeek-V4-Flash-DSpark structure. It integrates a speculative decoding module directly into the main checkpoint, boosting generation throughput without requiring an external draft model.

Smaller footprint, bigger agent punch

Despite having a far smaller activated parameter count, the post-trained V4-Flash consistently outperforms the larger DeepSeek-V4-Pro preview across standard benchmarks.

On Terminal Bench 2.1, it clocks 82.7, jumping past the V4-Flash preview's 61.8 and V4-Pro preview's 72.1, while edging out GLM-5.2 at 81.0 and trailing close behind Opus 4.8 at 85.0.

On DeepSWE, it hits 54.4 compared to V4-Pro preview's 12.8. In Toolathlon-Verified, it scores 70.3 against GLM-5.2's 59.9.

Community reviewers highlight that the model is roughly 70% smaller than GLM-5.2, while output token pricing is estimated to be up to 89 times cheaper than Claude Opus 4.8. It also includes out-of-the-box system prompt adaptations for Codex environments, enabling one-line drop-in configuration.

How to run it and requirements

The full unquantized weights on Hugging Face measure about 167GB, licensed under MIT. Unsloth has already rolled out GGUF quants, shrinking the smallest 1-bit profile down to around 82.5GB.

Both vLLM and SGLang natively support the release. You can turn on DSpark acceleration with a single speculative config flag, with draft weights pulled directly from the primary checkpoint.

For enterprise clusters, official recipes demonstrate serving across a single 4-way GB300 node using FP8 KV cache and FP4 indexing cache.

Keep in mind that DeepSeek omitted the standard Jinja chat template this time. Instead, developers need to use the provided Python scripts in the encoding repository to wrap OpenAI-compatible prompts. The reasoning_effort parameter now supports low, high, and max modes, with recommended max output lengths of up to 384K tokens.

Why it matters

If you build local code agents, command-line automation, or multi-step tool workflows, this is an immediate download. You get near-frontier agent precision without paying closed-source cloud bills.

We think DeepSeek just set a harsh standard for cost efficiency. By proving that focused post-training and integrated speculative draft layers can make a compact model punch alongside flagship models like GLM-5.2 and Opus 4.8, it makes running serious local agents genuinely affordable.