DeepSeek's 13B-Active Model Just Beat Its Own Flagship

TL;DR

On July 31, DeepSeek published DeepSeek-V4-Flash-0731, the official release of V4-Flash that supersedes the April preview. It is the same architecture and the same size: 284B total parameters, 13B active per token, MIT licensed, ungated, a million tokens of context. What changed is the post-training. On Terminal-Bench 2.1 the score went from 61.8 to 82.7, which is not just a big jump, it is past DeepSeek V4-Pro, the company's own much larger flagship, at 72.1. The cheap model ate the expensive one, and nobody had to buy more GPUs.

Nothing New Was Built

This is the part worth sitting with. DeepSeek did not widen the model, add experts, or extend the context. The model card is explicit that 0731 supersedes the preview with "substantially enhanced agentic capabilities," and every write-up that has looked at the config agrees the architecture and parameter count are untouched. The gains come from re-post-training.

If you have never had to explain the difference to a non-practitioner: pretraining is what the model absorbed about the world, and post-training is the part that teaches it how to behave on the job. DeepSeek did not send this model back to school. It reflashed the ECU on an engine block it already had, and the same block now makes considerably more power.

That distinction matters commercially, because re-post-training is cheap relative to a pretraining run. Whatever recipe produced this jump is repeatable, and DeepSeek can presumably point it at V4-Pro next.

The Agentic Numbers Are Not Subtle

DeepSeek's comparison table puts 0731 against both the V4-Flash preview and V4-Pro preview, and 0731 takes every row. The spread is what makes it interesting: the ordinary coding evals moved a normal amount, and the long-horizon agentic ones moved absurdly.

• DeepSWE: 7.3 to 54.4. The preview was, functionally, not able to do this task at all.
• CyberGym: 38.7 to 76.7, on a suite of 1,507 real-world C/C++ vulnerabilities where the agent has to write a proof-of-concept that actually reproduces the bug.
• Toolathlon-Verified: 49.7 to 70.3.
• NL2Repo: 39.4 to 54.2.
• Agents' Last Exam: 15.8 to 25.2.
• AutomationBench Public: 10.8 to 25.1.

A single-digit-to-54 move on DeepSWE is the tell. Scores that low usually mean the model was failing the harness rather than failing the problem: dropping tool calls, losing the thread across turns, giving up early. Fixing that is a post-training problem, not a capacity problem, which is exactly the kind of thing a fresh post-training run repairs.

The Flagship Problem

DeepSeek's own model card says the quiet part out loud: 0731 "outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count." That is a strange sentence for a company to publish about its two products. Somewhere in Hangzhou, the V4-Pro team is having a quiet week.

It is not a clean sweep against the frontier. Anthropic's Claude Opus 4.8 still leads Terminal-Bench 2.1 at 85.0, and on DeepSeek's internal DSBench sets Opus 4.8 posts 71.6 (FullStack) and 71.7 (Hard) against 0731's 68.7 and 59.6. But GLM-5.2 at 81.0 is now behind it on Terminal-Bench, and the gap to Opus is 2.3 points on a benchmark of 89 hard terminal tasks.

The Price Is the Actual Story

API pricing is $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens, with a 2,500 concurrency limit. Output supports up to 384K tokens at high or max reasoning effort, and the model exposes three effort levels: low, high, and max.

Independent evaluation backs the vibe. Artificial Analysis scores V4-Flash-0731 at 50 on its Intelligence Index, a 10-point jump over the April build, tying Google's Gemini 3.6 Flash and landing within a single point of GLM-5.2, Meta's Muse Spark 1.1, and GPT-5.6 Luna, all at 51. Artificial Analysis puts its cost per task at roughly 60% below Luna's for essentially the same measured intelligence, helped by a cache-hit discount of about 98% where the industry norm is closer to 90%.

On the intelligence-versus-cost chart, the new build sits almost directly above the old one at about the same price, which is the rare shape you want: a near-vertical move on the Pareto frontier rather than the usual diagonal drag.

Running It

Weights are on Hugging Face under MIT, ungated, with an attached DSpark speculative decoding module, the same family as the speed stack DeepSeek open-sourced earlier. Deployment guidance points at vLLM with DSpark enabled. Note the 304B figure floating around: that is the repo total including the draft module, while the model itself is 284B with 13B active.

Sampling recommendation from the card: temperature 1.0, with top_p 0.95 for agentic scenarios and 1.0 otherwise. There is no Jinja chat template, so the release ships dedicated encoding scripts instead. Unsloth has already mirrored it if you want quantized variants.

Caveats Worth Holding

Every benchmark number above except the Artificial Analysis index is self-reported by DeepSeek. Third-party replication on Terminal-Bench and CyberGym has not landed yet, and self-reported agentic scores are unusually harness-sensitive, which cuts both ways.

There is also a real cost the index numbers hide. Artificial Analysis logged 210 million tokens across its evaluation run, notably more verbose than comparable models. Cheap per token plus very long reasoning chains is not automatically cheap per task, so measure your own workload before assuming the price sheet. Simon Willison's first pass makes the same point from the other direction: default settings gave him subpar results, and raising reasoning effort is what unlocked the quality.

Key Takeaways

• Same model, retrained. V4-Flash-0731 keeps the 284B/13B-active architecture of the April preview. The entire improvement comes from post-training.
• It beats DeepSeek's own flagship. 82.7 on Terminal-Bench 2.1 against V4-Pro preview's 72.1, at a fraction of the activated parameters, and DeepSeek says so on the model card.
• Agentic evals moved hardest. DeepSWE 7.3 to 54.4 and CyberGym 38.7 to 76.7 suggest the preview was failing the loop, not the reasoning.
• $0.14 in, $0.28 out, MIT, ungated. With a 98% cache-hit discount, Artificial Analysis puts cost per task at roughly 60% below GPT-5.6 Luna at effectively equal measured intelligence.
• Verify before you bet on it. The benchmark table is self-reported, the model is verbose, and quality depends heavily on the reasoning-effort setting you pick.

Sources: DeepSeek-V4-Flash-0731 model card, Artificial Analysis, MarkTechPost, OfficeChai, OfficeChai on the Intelligence Index, Simon Willison, Terminal-Bench 2.1, CyberGym