DeepSeek V4-Flash-0731 Just Beat V4-Pro on Agent Benchmarks. Read That Again.

DeepSeek shipped a tiny re-post-train of V4-Flash today that, on its own dspark-speculative-decoding stack, beats the V4-Pro Preview on every agentic benchmark by margins that should make every model lab uncomfortable — including a 4.2x jump on DeepSWE.

DeepSeek shipped a quiet update today that buries one of the most annoying assumptions in the LLM space: that larger models are strictly better at agent work. The new V4-Flash-0731 — same architecture as the April preview, no new weights trickery, just better post-training — outperforms the V4-Pro Preview on every agent benchmark DeepSeek published, and most of them by embarrassing margins.

Let me put the headline numbers on the table before you scroll past:

Terminal Bench 2.1: 82.7 (vs V4-Pro Preview 72.1)
NL2Repo: 54.2 (vs V4-Pro Preview 38.5)
Cybergym: 76.7 (vs V4-Pro Preview 52.7)
DeepSWE: 54.4 (vs V4-Pro Preview 12.8)
Toolathlon Verified: 70.3 (vs V4-Pro Preview 55.9)
AutomationBench Public: 25.1 (vs V4-Pro Preview 12.8)
DSBench-FullStack: 68.7 (vs V4-Pro Preview 41.8)

On DeepSWE specifically — a software engineering agent benchmark — V4-Flash scored 54.4 against V4-Pro Preview's 12.8. That's not a marginal improvement. That's a 4.2x jump, from a smaller "Flash" model against the Pro-tier sibling. The same architecture, the same parameter count, the same inference cost class. Just better data, better RLHF, better post-training.

If you were charging three times the price for "Pro" thinking it bought you agent capability, today is a rough day at the office.

What Actually Changed

To be clear about what DeepSeek did and didn't do: V4-Flash-0731 is the same model as V4-Flash Preview under the hood. The architecture is unchanged. The parameter count is unchanged. The "activated" parameter count — what matters for MoE inference cost — is unchanged. The only thing that moved is post-training.

This is the part that should make every model lab sit down and shut up for a minute. The 2026 paradigm assumption has been that capability scales with parameters, and that the gap between a "Flash" and a "Pro" tier is mostly architectural. DeepSeek just demonstrated that the gap, for agent workloads specifically, is mostly data and post-training. The model size ceiling matters less than we thought.

The HF model card also ships a new component: DSpark, a speculative decoding module that DeepSeek cleverly attached to the Flash variant. With vLLM, you enable it with a single flag:

bash

--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

That gives you 7 speculative tokens per forward pass, with the draft model using greedy sampling to avoid wasted compute. The whole feature is documented in a single config line, which is the right amount of friction for a developer to actually bother enabling it.

The Reasoning Effort Knob

V4-Flash-0731 introduces a three-level reasoning_effort parameter: low, high, and max. This isn't a new idea — GPT-5.6 has it, Gemini 3.6 has it, Claude 4.8 has it — but DeepSeek's implementation is cleaner than most. The model is the same weights; only the post-training inference budget changes. You can route the same request through three different "thinking" budgets at runtime, and the cheaper the request, the less you pay.

This is the direction the industry is going and the right one. The "single model, single thinking mode" approach is dead. Different tasks want different inference budgets, and forcing everyone to pay for max effort on every call is a billing antipattern. DeepSeek joining the reasoning_effort crowd is just the market saying loud and clear that customers want this knob.

The Native Codex / Responses API Thing

The other note worth flagging: V4-Flash natively supports the Responses API format and is "specifically adapted for Codex." If you've been writing tool-calling agents over the past six months, you've probably hit the OpenAI Responses API / Anthropic Messages API / Chat Completions API / custom JSON format mess. DeepSeek adopting Responses API as the canonical Agent format is a quiet but significant move. It signals that the OpenAI wire format is winning the agent developer experience war, and labs that want to be where the developer tooling is going are aligning to it.

For Codex users specifically, this means DeepSeek is now a clean drop-in for coding agent backends. The codepath the harness exercises is the one the model was tested on. Fewer surprises, fewer integration bugs.

Where V4-Flash Lands

Compared to the rest of the field, here's the picture:

Terminal Bench 2.1: 82.7. Opus-4.8 scores 85.0. GLM-5.2 scores 81.0. V4-Flash is between the two flagship proprietary models on terminal tasks.
NL2Repo: 54.2. Opus-4.8 is at 69.7, but nobody else is close. V4-Flash is second here.
Cybergym: 76.7. Opus-4.8 is at 83.1. V4-Flash is second.
DeepSWE: 54.4. Opus-4.8 is 58.0. Essentially tied with the SOTA.
DSBench-Hard: 59.6. Opus-4.8 is at 71.7. Still chasing but competitive.

The shape of the curve is what matters. V4-Flash, MIT-licensed, open-weights, the cheap option — is within single-digit points of Opus-4.8 on agent benchmarks. The cheap model is no longer the bad model. It's the late model. And the gap is closing.

The Real Story

The DeepSeek V4-Flash-0731 release is happening because DeepSeek is about to ship V4-Pro. The official Pro release is "following soon" according to their changelog. V4-Flash is the preview of what the Pro tier will do — and on agent benchmarks, the Flash model is already eating the Pro tier's lunch. Imagine what the actual Pro release will look like.

The bigger lesson, the one DeepSeek is teaching us in public: in 2026, agent capability is not a function of model size. It's a function of post-training quality. The Flash model is what you ship when you want to be cheap. The Pro model is what you ship when you want to be the best. But "cheap" doesn't mean "worse" anymore. It means "more efficient inference for very nearly the same capability."

That's the death of the "scale at all costs" assumption. DeepSeek's MiniMax-Hailuo-02 competitors are going to have to come up with better post-training recipes, not just bigger parameter counts, if they want to keep the flagship crown. The race is no longer about who can train the largest model. It's about who can post-train the best agent.

Expected it. Welcome to the new normal.