DeepSeek V4 Flash Benchmarks: Speed and Scores Decoded

If you searched "DeepSeek V4 Flash benchmarks," you probably want two numbers fast: how many tokens per second it pushes, and how it scores on the coding and reasoning suites everyone quotes. The problem is that launch-week benchmark charts are the noisiest data in the industry. Vendors pick the tests they win, third-party aggregators list wildly different throughput depending on the host, and half the "scores" circulating on social feeds have no primary source at all.

This article does the boring but useful thing: it separates what is actually verifiable about DeepSeek V4 Flash from what is not. Everything stated as a hard number here traces back to DeepSeek's own documentation or an independent measurement authority (primarily Artificial Analysis). Where a specific benchmark figure could not be confirmed against a primary source, we say so plainly and describe it qualitatively instead of inventing a value. This is written as of August 2026; model routing, prices, and leaderboard positions move quickly, so treat every figure as a dated snapshot and re-check the linked sources before you budget against them.

What "Benchmarks" Mean for a Flash-Tier Model

DeepSeek V4 Flash is the efficiency tier of the DeepSeek V4 family. Its sibling, DeepSeek V4 Pro, is the large flagship (a roughly 1.6-trillion-parameter mixture-of-experts model built to win the hardest reasoning tests). Flash plays a different game. It is a Mixture-of-Experts model with 284B total parameters and only ~13B activated per token, designed for speed, low cost, and high-volume "everyday work" rather than topping a reasoning leaderboard.

That design goal changes which benchmarks matter. For a Flash-class model, the numbers worth reading are:

• Throughput (output tokens per second) and latency (time to first token) — the reason you would pick Flash over Pro in the first place.
• Cost per million tokens — Flash exists to be cheap at scale.
• Task-level quality scores (coding, agentic tool use, MMLU-style knowledge) — but read as "is it good enough for this job," not "does it beat the frontier."

A model that is twice as fast and five times cheaper does not need to win MMLU. It needs to be accurate enough that the speed and price become the deciding factor. So we will start with what is verified, then be explicit about what is not.

The Verified Numbers

These are the specifications you can confirm today from DeepSeek's official documentation and independent model registries. This is the only table in this article filled entirely with sourced figures.

Two things stand out. First, the 1M-token context is genuinely flagship-class — Flash does not cut context to hit its price point, which is unusual for an efficiency tier. Second, at $0.14 input / $0.28 output, it is priced to be run millions of times, not admired once. (Some third-party hosts, such as DeepInfra, list even lower rates around $0.10 / $0.20 — those are provider-specific and should not be read as DeepSeek's own pricing.)

Throughput and Latency: What the Numbers Say

This is where "benchmarks" get slippery. Output speed for a hosted model is not an intrinsic property of the weights — it depends on the serving stack, batch size, quantization, and which provider you hit. There is no single "true" tokens-per-second for DeepSeek V4 Flash, which is why we are not quoting one.

What is defensible: the ~13B active-parameter design is the entire reason Flash is fast. Activating a small slice of a 284B MoE per token is the standard architectural lever for high throughput and low latency, and it is why DeepSeek positions Flash as the low-latency, high-volume option versus the heavier V4 Pro. Independent trackers like Artificial Analysis publish live median output-speed and time-to-first-token measurements per provider; as reported by those measurement pages, Flash sits in the fast, low-latency tier rather than the slow-but-smart tier. For an actual number to plan against, read the current figure on the provider you intend to use — a single screenshot from launch week is not a spec.

Honest flag: any specific "X tokens/sec" or "Y ms latency" you see attributed to DeepSeek V4 Flash is provider- and date-specific. We could not tie a single canonical throughput number to a primary source, so we describe it qualitatively rather than print a figure that would be wrong for most readers' setups.

Published Benchmark Scores: What We Can and Can't Confirm

Here is the part most "DeepSeek V4 Flash benchmarks" pages get wrong — they list MMLU, coding, and agentic scores with two-decimal precision and no citation.

Being straight with you: we could not confirm specific, primary-sourced benchmark scores (MMLU, SWE-bench-style coding, or agentic tool-use suites) for the Flash variant specifically at the time of writing. Many circulating numbers conflate Flash with V4 Pro, or come from unlabeled reposts. Per our sourcing rule, we will not print a score we cannot trace to DeepSeek or an independent authority.

What can be said qualitatively, and how to verify it yourself:

• Coding and agentic tasks: DeepSeek markets Flash for coding assistants and agentic workflows, and its non-reasoning-by-default profile favors fast, cheap iteration over deep chain-of-thought. As reported by aggregate leaderboards, the Flash tier trades some peak reasoning quality for speed — expected for its size. Confirm the current standing on the Artificial Analysis Intelligence Index page for the model.
• Knowledge (MMLU-style): treat any single MMLU figure for Flash as unverified unless it links to a primary run. A ~13B-active model is capable but is not built to top graduate-level knowledge tests.
• The right mental model: Flash is an "is it good enough, and is it fast and cheap?" model. If a task needs the highest possible score, that is a job for a flagship (V4 Pro, or GLM 5.2 — see below), not for Flash.

If you need real scores, the reliable path is to open the model's page on an independent benchmark authority and read today's number, then run your own prompts. Do not trust a chart with no link.

DeepSeek V4 Flash vs GLM 5.2: A Verified-Spec Comparison

This is the decision that matters for most readers: cheap-and-fast versus flagship depth. The table below compares only verified specifications — not contested benchmark scores — so you can reason about the trade-off without relying on numbers we cannot stand behind.

Read it honestly. DeepSeek V4 Flash is the clear winner on raw cost and speed — roughly 10x cheaper on input than GLM 5.2, with a lighter active footprint built for throughput. GLM 5.2 activates about 3x more parameters per token and is the stronger flagship for demanding coding and agentic work, where a benchmark point or a correctly resolved bug is worth far more than a fraction of a cent. Both carry a full 1M-token context, so this is not a context trade-off — it is a quality-versus-cost trade-off.

The practical routing rule: reach for DeepSeek V4 Flash when volume and latency dominate, and for a flagship like GLM 5.2 when the depth of the answer is the thing you are paying for. They are not really competitors; they are two different budget lines.

The fastest way to feel that difference is to run the same real task through both. You can open GLM 5.2 in your browser on glm5.app with no API key and no install, paste a real ticket or diff, and judge the flagship-tier output directly against a Flash-tier run. A chart is an average over someone else's tests; your own prompt is the only benchmark that ships.

How to Read Flash Benchmarks Before You Commit

Because so many DeepSeek V4 Flash "benchmark" claims are unsourced, here is the judgement checklist we used for this article, and that you can reuse:

1. Does the number link to a primary source? DeepSeek's docs for specs and pricing; an independent authority (Artificial Analysis) for live speed and index scores. No link, no trust.
2. Is it about Flash specifically, or V4 Pro? The two are routinely conflated. A "V4" score is often the Pro number.
3. Is the throughput figure tied to a provider and date? Speed is a serving-stack property, not a weights property.
4. Does the benchmark match your job? A coding score is meaningless for a summarization pipeline, and vice versa.
5. Did you run your own eval? Five real prompts from your workflow beat any leaderboard for a routing decision.

Frequently Asked Questions

What are the confirmed specs behind DeepSeek V4 Flash benchmarks? Verified from official/registry sources: a 284B-total, ~13B-active MoE model, 1M-token context, up to 384K max output, priced at $0.14 input / $0.28 output per million tokens, MIT-licensed, released 2026-04-24.

How fast is DeepSeek V4 Flash in tokens per second? There is no single canonical figure — output speed depends on the provider and serving stack. Its ~13B active-parameter design places it in the fast, low-latency tier; check the live median on your chosen provider's Artificial Analysis page for a current number.

What is DeepSeek V4 Flash's MMLU or coding score? We could not confirm a primary-sourced MMLU or coding score for the Flash variant specifically, and will not print an unverified one. Read the current figure on an independent benchmark authority and validate with your own prompts.

Is DeepSeek V4 Flash better than GLM 5.2? Different tiers. Flash wins on cost and speed (roughly 10x cheaper input). GLM 5.2 is the flagship for deep coding and agentic quality. Pick by whether volume or answer-depth dominates your task.

Is DeepSeek V4 Flash open-weight? Yes — MIT-licensed and open-weight, so you can self-host it, though most users access it through a hosted API.

The Bottom Line

The verifiable story on DeepSeek V4 Flash is strong and simple: a 284B/~13B-active MoE with a full 1M-token context, priced to run at scale, built for speed over peak scores. The unverifiable story — the specific tokens-per-second and MMLU/coding numbers floating around — deserves more skepticism than it usually gets, which is why we flagged rather than fabricated it.

If your workload is high-volume and cost-sensitive, Flash is a rational default. If you are paying for the depth of the answer — hard coding, long agentic chains — that is flagship territory. Test a flagship the same way you would test Flash: with your own task, right now. Try GLM 5.2 free on glm5.app, run your real prompt, and see the pricing at glm5.app/pricing if you decide to build on it.

Author: GLM 5 Team. Reviewed against DeepSeek official documentation and Artificial Analysis as of August 2026. Specs, prices, and benchmark standings change frequently — verify the linked sources before quoting any figure.

Sources

• DeepSeek Models & Pricing — Official model IDs, context length, 384K max output, and token pricing for deepseek-v4-flash.
• DeepSeek Your First API Call — Official base URL, supported V4 model IDs, and OpenAI-format request examples.
• DeepSeek Chat Completions API — Official request schema and supported parameters.
• DeepSeek Thinking Mode — Official reasoning-effort behavior, relevant to Flash's non-reasoning-by-default profile.
• DeepSeek platform — Official account, billing, and API-key management.
• DeepSeek on Hugging Face — DeepSeek's official open-weight model releases and licenses.
• Artificial Analysis — DeepSeek V4 Flash — Independent measurements of output speed, latency, and index scores per provider.
• Artificial Analysis — Methodology — How the independent throughput, latency, and Intelligence Index figures are measured.
• OpenRouter — DeepSeek V4 Flash — Independent model registry with parameter, context, and provider pricing metadata.
• GLM 5.2 on glm5.app — Zhipu AI's flagship GLM 5.2, used here for the verified-spec flagship comparison.