DeepSeek V4 Flash 0731 Benchmarks: 13B Active Beats 1.6T

The DeepSeek V4 Flash 0731 benchmarks landed on July 31, 2026 with a detail almost every writeup skipped: DeepSeek did not change a single thing about the model’s architecture. Same structure, same parameter count — and roughly double the agent scores anyway. On one test, more than seven-fold.

DeepSeek-V4-Flash-0731 has the same 284B-parameter Mixture-of-Experts structure as the preview build it replaces. Same total parameters. Same ~13B active per token. Same 1M context window. DeepSeek’s own model card says it “keeps the same model architecture and size” and was “only re-post-trained.”

That one sentence is the most important thing in the release, and almost every writeup buried it under a leaderboard screenshot. If you care about where agentic capability actually comes from, this is the cleanest natural experiment the open-weights world has handed us all year.

TL;DR

• The architecture didn’t change. V4-Flash-0731 is the preview build with a new post-training run — 284B total / ~13B active MoE, 1M context, MIT license, published openly on Hugging Face.
• Agent scores moved violently. Terminal Bench 2.1 went 61.8 → 82.7. DeepSWE went 7.3 → 54.4. That is a 7.5x jump from post-training alone.
• It beats V4-Pro Preview on all nine published agent benchmarks while activating roughly a quarter of Pro’s parameters per token.
• Price is the real story for engineers: $0.14 / $0.28 per million tokens versus Pro’s $0.435 / $0.87 — about 3.1x cheaper on both sides.
• The caveat you must hold onto: the comparison target is Pro Preview, and every number comes from DeepSeek’s own harness. Benchmark it yourself before you migrate anything.

What is DeepSeek V4 Flash 0731?

DeepSeek-V4-Flash-0731 is the official, non-preview release of DeepSeek’s efficiency-tier V4 model: a 284-billion-parameter Mixture-of-Experts system that activates about 13 billion parameters per token, handles a 1-million-token context window, ships with a speculative decoding module attached, and is licensed MIT for commercial use and self-hosting.

The naming is a date stamp — 0731 is July 31, 2026 — which is DeepSeek’s way of saying this is a build, not a new generation. It supersedes V4-Flash-Preview and sits below V4-Pro, the 1.6-trillion-parameter flagship that activates roughly 49B per token.

Text in, text out. No vision, no audio, no video. That is a deliberate scope choice, and it is part of why the price is what it is.

The DeepSeek V4 Flash 0731 benchmarks, in full

Here is what DeepSeek published, comparing the new build against its own preview and against the Pro preview:

Benchmark	V4-Flash-0731	V4-Flash (Preview)	V4-Pro (Preview)
Terminal Bench 2.1	82.7	61.8	72.1
NL2Repo	54.2	39.4	38.5
Cybergym	76.7	38.7	52.7
DeepSWE	54.4	7.3	12.8
Toolathlon-Verified	70.3	49.7	55.9
Agents’ Last Exam	25.2	15.8	16.5
AutomationBench Public	25.1	10.8	12.8
DSBench-FullStack	68.7	37.0	41.8
DSBench-Hard	59.6	25.8	31.1

Nine benchmarks. Nine wins over both comparison points. The smallest relative gain is Terminal Bench at +34% over the preview; the largest is DeepSWE at roughly 7.5x.

Independently, Artificial Analysis scores the model at 50 on its Intelligence Index, well above the median for open-weight models in its size class — though it also flags the model as unusually verbose, generating 210M tokens across the evaluation suite against a 100M median. Verbosity is not free when you pay per output token.

💡 Key insight: When a model’s architecture is frozen and only post-training changes, every point of benchmark movement is attributable to post-training. This release is a controlled experiment that most labs never publish.

Why post-training moved the needle this hard

Pretraining gives a model knowledge. Post-training gives it behavior — and agentic work is almost entirely behavior.

Think about what DeepSWE and Terminal Bench actually measure. Not “does the model know what a race condition is.” They measure whether it will run a command, read the output honestly, notice it was wrong, back out, and try a different approach — twenty times in a row without losing the thread. That is a policy problem, not a knowledge problem.

A 7.3 on DeepSWE doesn’t mean the preview build lacked the knowledge to fix bugs. It means it fell out of the loop: gave up early, hallucinated a passing test, or burned its context re-reading files it had already read. Those are all correctable with better post-training, and DeepSeek corrected them.

This is the same lesson that keeps showing up everywhere agents are deployed. The large-scale study of AGENTS.md files found that what changed agent outcomes most wasn’t raw model capability — it was the scaffolding and instructions shaping how the model behaved inside a loop. DeepSeek just demonstrated the same thing from the model side of the boundary.

The uncomfortable implication for anyone budgeting compute: the agentic gap between models is now mostly a post-training gap, and post-training is dramatically cheaper than pretraining. Parameter count is losing its predictive power for the tasks engineers actually care about.

Where the price math lands

Benchmarks pick the shortlist. Price picks the model.

	V4-Flash	V4-Pro
Input / 1M tokens	$0.14	$0.435
Output / 1M tokens	$0.28	$0.87
Total params	284B	1.6T
Active params / token	~13B	~49B
Context	1M	1M

Flash is about 3.1x cheaper on both input and output, per DeepSeek’s published V4 pricing.

Make it concrete. A coding agent that consumes 20M input tokens and produces 2M output tokens in a day — a plausible load for a small team running agents against a real repository — costs $2.80 + $0.56 = $3.36/day on Flash, versus $8.70 + $1.74 = $10.44/day on Pro. Over a year that’s roughly $1,225 against $3,810.

That gap is what makes the benchmark result interesting rather than academic. A model that is both cheaper and better on your workload isn’t a tradeoff; it’s a decision. The question is only whether “better on the published benchmarks” survives contact with your codebase.

Four ways people are misreading this release

1. “Flash beats Pro.” It beats Pro Preview. DeepSeek explicitly re-post-trained Flash and compared against an older Pro build. When Pro gets the same treatment, expect the ordering to change. Reading this as “small models won” is exactly the wrong takeaway.

2. Treating vendor benchmarks as harness-independent. The model card notes these are code-agent tasks run with DeepSeek’s own harness at specific reasoning-effort levels. Your harness is not their harness. Numbers this good are a reason to run an evaluation, not a substitute for one.

3. Assuming 13B active means 13B of hardware. All 284B parameters have to be resident to route between experts. The active count buys you inference speed, not memory relief. If you’re planning to self-host, the constraint is the same one that shapes every local LLM coding setup: VRAM first, everything else second.

4. Ignoring the verbosity tax. Artificial Analysis measured this model generating more than twice the median token count. On a per-output-token bill, a chatty model erodes its own price advantage. Measure cost-per-completed-task, not cost-per-million-tokens.

What to actually do this week

1. Pull the weights or hit the API and run your own agent harness against three or four real tickets from your backlog. Not SWE-bench. Your repository, your test suite, your CI.
2. Measure cost per resolved task, not per token. That is the only number that survives the verbosity issue.
3. Watch the context degradation curve. A 1M-token window is a capacity claim, not a quality claim. Find where retrieval accuracy actually falls off for your codebase.
4. Keep a second vendor wired up. Open-weight releases move fast and pricing moves faster; the teams that win are the ones who can swap models in an afternoon.
5. Don’t skip the governance question. Open weights don’t remove data-handling obligations if you use the hosted API, and the broader distillation and provenance fight around Chinese open-weight releases is still live.

The honest limitations

Three things I can’t tell you yet, and neither can anyone else on day two.

Independent agent benchmarks haven’t caught up. Every number in that table is DeepSeek’s. The Artificial Analysis Intelligence Index of 50 is independent, but it’s a general-capability score, not an agentic one, and it isn’t measuring the thing that jumped.

Long-horizon reliability is unproven. Benchmarks run bounded tasks. Real agent work runs for hours and fails in ways that only show up at hour three.

And the throughput picture is incomplete — output tokens per second and time-to-first-token weren’t disclosed at launch. The speculative decoding module suggests DeepSeek is optimizing hard for latency, but “suggests” is not “measured.”

FAQ

What is DeepSeek V4 Flash 0731?

DeepSeek-V4-Flash-0731 is the official release of DeepSeek’s efficiency-tier V4 model, published on July 31, 2026 under an MIT license. It is a Mixture-of-Experts model with 284 billion total parameters and roughly 13 billion active per token, a 1M-token context window, and an attached speculative decoding module. It supersedes the earlier V4-Flash preview build.

How is it different from the preview version?

The architecture and parameter count are identical — DeepSeek states the model was only re-post-trained. Everything that changed is downstream of pretraining: agent scaffolding, tool-use behavior, and long-horizon task discipline. That single post-training cycle moved Terminal Bench 2.1 from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4.

Does DeepSeek V4 Flash really beat V4-Pro?

It beats V4-Pro Preview on all nine agent benchmarks DeepSeek published, despite activating roughly 13B parameters against Pro’s ~49B. That is a real result, but the comparison target is a preview build, not a finished flagship. Expect the gap to narrow or invert once Pro gets the same post-training treatment.

How much does it cost compared to V4-Pro?

V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens. V4-Pro is $0.435 input and $0.87 output. That makes Flash roughly 3.1x cheaper on both sides of the ledger, which is the number that actually decides model selection for high-volume agent loops.

Can I self-host DeepSeek V4 Flash 0731?

Yes. The weights are published in a public, ungated Hugging Face repository under the MIT license, which permits commercial use and self-hosting. The practical constraint is memory, not licensing — 284B total parameters must stay resident even though only ~13B are active per token.

Should I switch my coding agent to it?

Not on the benchmark numbers alone. Run your own harness against your own repositories for a week before moving anything, because the published agent scores were measured with DeepSeek’s harness and reasoning-effort settings. Treat this release as a strong reason to run that evaluation, not as a result you can inherit.

Final take

The headline isn’t that DeepSeek shipped a cheap model that scores well. It’s that they proved, with the architecture held constant, that agentic capability is a post-training artifact.

That reframes the next year. If a single post-training cycle can take a model from 7.3 to 54.4 on DeepSWE, then the frontier for agent work isn’t going to be decided by who has the most parameters — it’ll be decided by who is best at teaching a model how to behave inside a loop. That’s a much more contestable race, and open-weight labs are unusually well positioned to run it.

Five months ago the story around this model family was geopolitics and speculation. Now there are weights on Hugging Face, a price list, and a benchmark table. Go measure it against your own work — and if you’re deciding which agent to point at it, start with how the major coding agents actually differ.

Sources

• DeepSeek-V4-Flash-0731 model card — Hugging Face. Benchmark table, architecture notes, MIT license.
• DeepSeek V4 Flash intelligence, performance and price analysis — Artificial Analysis. Intelligence Index score, verbosity measurement, pricing.
• DeepSeek V4: V4-Pro and V4-Flash complete guide — DeepSeek. Parameter counts, context window, official API pricing.

---

Written for umesh-malik.com — no-fluff technical writing on AI, Web Dev, and Engineering.