DeepSeek V4-Flash [What It Changes]

TL;DR

• DeepSeek V4-Flash is DeepSeek's cheaper open-weight model that, as of July 31, 2026, can use tools, run code, and finish multi-step jobs with less hand-holding than its April preview.

• The company kept the same architecture and size as the preview; the official jump came from more training, and the weights stay free to download under MIT.

• Tests after launch found official Flash beating DeepSeek's own pricier V4-Pro preview on several agent tasks, which is why a two-truck HVAC shop or a 10-person agency can treat Flash as the default worker.

• API list prices later moved to peak and off-peak bands; Flash remains the low-cost line, while V4-Pro is the paid step-up for harder agent work.

Key Takeaways

• Use DeepSeek V4-Flash when the job is a tool-using agent — inbox triage, form fill, code edits — and you want a model swap, not a rebuild.

• Call the API as deepseek-v4-flash; the July 31 build is tagged DeepSeek-V4-Flash-0731 and still speaks OpenAI and Anthropic request formats.

• Keep V4-Pro for the hardest coding-agent runs; Flash closed most of the April gap, then Pro's August build pulled ahead again on the same tests.

• Schedule heavy batch jobs in off-peak hours if you pay DeepSeek directly: off-peak rates are half of peak.

• If privacy is the constraint, the same Flash checkpoint is on Hugging Face under the MIT License, so a clinic can host it instead of sending tokens out.

DeepSeek V4-Flash is the official, freely downloadable DeepSeek model that turns cheap inference into a practical office agent: it can call tools, run code, and carry a multi-step task farther than the April preview without a new architecture.

That sentence is the whole product for a small shop. A two-truck HVAC company does not need a 1.6-trillion-parameter flagship to turn yesterday's invoices into today's job board. A 10-person marketing agency does not need a $50-per-million-token closed model to draft a client recap, pull CRM fields, and file the PDF. A solo clinic does not need a research lab to extract intake forms and route them into the chart.

If your team already maps executive-assistant chores onto agents, the July 31 release is a model-id change. You keep the same workflow graph, point the model step at deepseek-v4-flash, and watch the bill and the error rate. That is the SMB bridge: China's AI race produced a cheaper worker for the same checklist a small U.S. office already runs.

What happened on July 31

According to Caixin Global, DeepSeek published the official DeepSeek-V4-Flash on July 31, 2026, with stronger autonomous-agent skills and another API cost cut relative to the path that led into that week. Caixin also reported the update landed a little after a mid-July target and without the V4-Pro general-availability build, which still sat in preview.

The company's own API changelog matches that date. The July 31 note says the official Flash API is in public beta, the caller keeps the model name deepseek-v4-flash, and DeepSeek-V4-Flash-0731 "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." They did not restack the engine. They kept practicing the same engine on agent work until the scores moved.

Preview Flash could already chat, reason, and take a 1-million-token context. It was weaker at the jobs that actually spend money in a shop: opening a terminal, writing a repo, calling a tool, finishing a ticket. The official build is the same size, aimed harder at those jobs.

DeepSeek's April preview note had already positioned Flash as the "fast, efficient, and economical choice" next to V4-Pro. The official Flash note then claimed the new scores "far exceed" the V4-Pro preview. That vendor line is why later write-ups treated Flash as the surprise, not Pro.

The weights are not a private binary. The DeepSeek-V4 collection on Hugging Face lists DeepSeek-V4-Flash-0731 as a public checkpoint, and the 0731 model card says the repository and weights use the MIT License. A shop that cannot send patient notes or tax files to a public API can still download the same official build.

The mechanism, without the equations

Flash is a mixture-of-experts model: a large pile of specialists, with only a slice turned on for each token. According to Caixin Global, the Flash line has 284 billion total parameters and 13 billion active parameters, against Pro's 1.6 trillion total and 49 billion active, and both sides share a 1-million-token context window. The V4 technical report restates those counts and adds that both models were pre-trained on more than 32 trillion tokens.

What a 10-person agency should hear: you pay for 13 billion active parameters, not 284 billion, on every token. That is why Flash can be fast and cheap while still swallowing a long PDF, a CRM export, and a thread of tool results in one window.

The April paper also describes the long-context trick in one sentence worth keeping. In a 1-million-token setting, V4-Pro needs only 27% of the single-token inference FLOPs and 10% of the KV cache versus DeepSeek-V3.2, per the arXiv abstract and the preview model card. Flash uses the same hybrid-attention family. The practical result is that "paste the whole job folder" stopped being a research demo.

Agent skill is not a mystery setting. In the tool-calls guide, the model does not run your functions. It names a function and arguments; your code runs the function and sends the result back; the model continues. Thinking mode can do several of those loops before it answers, as the thinking-mode guide spells out. If you send tools, you must pass the hidden reasoning text back on later turns or the API returns a 400.

The official Flash build also speaks the OpenAI Responses API format and is wired for Codex, per the July 31 changelog and the Responses API guide. A team that already drives coding agents through that format can keep the harness and change the model string. DeepSeek's first-call docs list the live names as deepseek-v4-flash, deepseek-v4-pro, and the experimental deepseek-v4-flash-vision-exp.

Thinking effort is now a three-way knob for both Flash and Pro: low, high, and max, with high as the default when thinking is on, per the thinking-mode table. Low is for cheap, short answers. High is the daily agent setting. Max is the "spend tokens until the hard ticket yields" setting. The 0731 card tells local operators to allow up to 384K output tokens at high and max effort.

None of that requires a new workflow product. Teams already routing documents through US Tech Automations workflows will plug this in as a model swap, not a rebuild.

What the scores actually say

The 0731 model card publishes a side-by-side that is the cleanest view of the July 31 jump. Official Flash is compared with Flash preview, Pro preview, GLM-5.2, and Opus-4.8 on the same agent suite. Official Flash scored 82.7 on Terminal Bench 2.1, against 61.8 for Flash preview and 72.1 for Pro preview, on that Hugging Face table.

Sources: DeepSeek-V4-Flash-0731 model card; matching figures also appear on the API changelog.

Read the DeepSWE row twice. Preview Flash sat at 7.3. Official Flash sits at 54.4. That is why developers talked about Flash as if Pro had been leapfrogged — because, on the April Pro preview, it had.

The same changelog later records the August 13 Pro general-availability scores. V4-Pro-0813 puts Terminal Bench 2.1 at 87.9, NL2Repo at 61.5, DeepSWE at 62.7, and Toolathlon-Verified at 74.1, per the August 13 changelog block. Flash is no longer ahead of current Pro. It is still close enough that a shop should default to Flash and overflow to Pro.

According to Reuters, independent firm Artificial Analysis gave reasoning V4-Pro a 53 on its Intelligence Index versus 40 for V4-Flash, on a nine-part mix of agent work, tools, coding, science, and long context. The live Artificial Analysis page for Flash 0731 (max effort) now lists that build at 52 on the same index family, with about 107.9 output tokens per second. Treat the Reuters 40 and the current 52 as dated snapshots, not a fight: effort level and calendar week both move the headline.

Global Times reported on August 4 that Artificial Analysis had Flash at 50, one point behind GPT-5.6 Luna at 51, while Flash output sat at $0.28 per million tokens against about $1.20 for Luna. That $0.28 figure is the pre-hike list, not the peak rate now on DeepSeek's price page.

On distribution, according to Global Times, DeepSeek V4 Flash processed 7.1 trillion tokens on OpenRouter in the week of July 27–August 2 and ranked first, while coding host OpenCode reported 8 trillion tokens in a single day and about a 30 percent rise in usage after the official launch. OpenRouter logged 7.1 trillion Flash tokens in one week, per that Global Times report. OpenRouter's Flash listing is the public catalog page for that routing layer; it shows a 1M context window and a wide spread of host prices.

Pricing: the cheap line, then the peak clock

DeepSeek did not keep a single forever price. In late May it made a steep cut on the flagship. In mid-August it raised V4 rates and split the day into peak and off-peak.

Caixin reported the May move in the open lede of its May 25 story: V4-Pro API pricing was cut 75 percent, and cache-miss input for Pro fell to 3 yuan ($0.44) per million tokens. Yahoo's Quartz write-up later noted a 75 percent promotional discount on V4-Pro that had run through May 5 before that permanent cut.

The live sheet is the Models & Pricing page. Flash peak output is $1.32 per million tokens on that DeepSeek table, with off-peak at $0.66. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; every other hour is off-peak, at half the peak rate. The same page sets Flash cache-miss input at $0.44 peak and $0.22 off-peak, cache-hit input at $0.014 peak and $0.007 off-peak, a 1M context, a 384K max output, and a 2500 concurrency cap, against 500 for Pro.

Source: DeepSeek Models & Pricing. Peak hours: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday.

According to Reuters, Artificial Analysis listed V4-Pro-0813 at $1.32 per million input tokens and $3.96 per million output tokens, about 9 times Flash input and 14 times Flash output at the Flash rates then listed as $0.14 and $0.28. Those Flash figures are the older flat list, not the current peak sheet. After the hike, the official table is a clean 3 times: $1.32 / $0.44 on cache-miss input, $3.96 / $1.32 on output.

The company said the new prices take effect at 16:00 UTC on August 16, 2026. A separate Reuters brief said the new rates range from 50% to 1,100% above then-current prices depending on model, token type, and time of use, and described the effective date as August 17. For a shop's calendar, use the vendor clock: 16:00 UTC, August 16, on the GA release note.

Yahoo put Pro output moving to $3.96 per million at peak from a $0.87 flat rate, and put Anthropic's Fable 5 at $50 per million output tokens. Even after the hike, Flash peak output at $1.32 is not in that neighborhood.

Global Times also printed yuan prices for Pro-0813 — 3 yuan ($0.44) per million input, 6 yuan per million output, 0.025 yuan per million cached input — and a 1-million-token context with 384,000-token max output. Use DeepSeek's dollar table for billing; the yuan quote and the context caps match the docs.

A clinic that already pushes form data into a CRM can keep that pipeline and change only the model behind the extract step. That is the price story in operational terms: you are no longer choosing between "too dumb" and "too expensive" for a nightly batch of intake PDFs.

USTA analysis: cost of 1 million peak output tokens

This is a derivation from figures already cited, not a new survey.

Inputs: DeepSeek list price for Flash peak output is $1.32 per 1 million tokens on the pricing page. DeepSeek list price for Pro peak output is $3.96 per 1 million tokens on the same page. Yahoo's Quartz story puts Anthropic Fable 5 at $50 per 1 million output tokens at this URL.

Arithmetic: $3.96 ÷ $1.32 = 3.00, so Pro peak output costs 3.00 times Flash. $50 ÷ $1.32 ≈ 37.88, so Fable 5 output costs about 37.9 times Flash at those list rates. Off-peak Flash output is $0.66, which is exactly half of $1.32, matching DeepSeek's "off-peak is half of peak" rule on the same page.

USTA analysis. Inputs: DeepSeek pricing; Yahoo / Quartz on Fable 5.

For a bookkeeping desk, that table is the decision. If a nightly close spends a few million output tokens of agent traces, Flash is the default. Pro is the overflow when a tax engine ticket fails. Pair that with practice-management software already in the stack rather than buying a new "AI close" product.

Timeline of the V4 family

Sources: changelog; April preview note; Caixin on the 75% cut; pricing page.

The retired names matter for old scripts. If a shop still calls deepseek-chat, that name went dark after July 24, 2026, 15:59 UTC, per the April preview note. Point those jobs at deepseek-v4-flash and set thinking on or off explicitly.

Flash versus Pro versus hosting it yourself

Use Flash when the task is high-volume and tool-shaped: inbox to CRM, PDF to ledger, repo cleanup, meeting notes to a task list. Use Pro when Flash fails the same ticket twice — the August Pro build is the one that scored 87.9 on Terminal Bench 2.1 in the changelog. Use a local 0731 checkpoint when the file cannot leave the building.

Self-host is real, not a slogan. The 0731 card ships vLLM and SGLang commands, including a speculative-decoding flag (DSpark) that rides in the same checkpoint. Hugging Face listed about 4.65 million downloads in the last month on that card. The older preview card at DeepSeek-V4-Flash is still up; the official swap is 0731. Pro weights live at DeepSeek-V4-Pro and the later 0813 listing.

Hardware is the honest limit. Flash is still a 284-billion-parameter MoE. A two-truck shop will call the API. A regional hospital IT team might host. Do not pretend a gaming PC is enough because the active count is 13 billion; the full expert pile still has to sit on disk and, for local serving, on accelerators.

The DeepSeek GitHub org is the other half of the agent story. The public deepseek-harness repo is listed there with about 208 thousand stars; the API landing page points harness developers at a separate quickstart. You do not need that harness to call Flash. You need it if you are rebuilding the evaluation loop DeepSeek used on Terminal Bench.

DeepSeek's homepage is the consumer door for the app and API. For an SMB, skip the chat window as the system of record. Put Flash behind the same small-business automation graph you already trust, with logs, retries, and a human on the last mile.

A tax shop comparing Drake, ProConnect, and UltraTax does not replace those engines with Flash. It uses Flash to classify organizers, extract K-1 fields, and draft client emails, then writes back into the engine it already paid for. A finance lead comparing Fathom and Jirav-style reporting can let Flash draft the commentary layer while the system of record stays put.

US Tech Automations sits in that same slot: the workflow is the product, and Flash is a replaceable worker inside it.

Limits you should budget for

Flash is not a magic intern. The 0731 table still trails Opus-4.8 on Terminal Bench 2.1 (82.7 vs 85.0), NL2Repo (54.2 vs 69.7), and DSBench-Hard (59.6 vs 71.7). AutomationBench Public is 25.1. Agents' Last Exam is 25.2. Those are not "solved." They are "cheap enough to retry."

Thinking mode ignores temperature and top_p; the docs say setting them will not error and will not work. Max output is 384K, not infinite. The Responses API is stateless: previous_response_id and conversation are not supported, per the compatibility table. If you need memory, you store it.

Vision is a different model id. deepseek-v4-flash-vision-exp is experimental, billed like Flash, and added on August 21 with its own scores in the changelog. Do not send screenshots to plain Flash and expect pixels to count.

Price risk is real. Reuters reported a $7.4 billion first outside round in June and talks of a further round around a $74 billion valuation in the August 13 feature. A company that just raised, and that already reversed a cheap promotional era with peak pricing, can move the sheet again. Cache-hit rates and off-peak clocks are how a shop keeps a bill predictable anyway.

Data-handling is a policy choice, not a benchmark. MIT weights let you stay on-prem. The public API does not. A clinic should decide that before the first PHI prompt, not after.

Signal vs Speculation

Signal (sourced): Official DeepSeek V4-Flash shipped July 31, 2026. Architecture and size match the April preview; gains are post-training. Weights are on Hugging Face under MIT. Vendor agent scores for 0731 beat the V4-Pro preview, then official Pro-0813 on August 13 reclaimed the lead on the same tests. The API speaks Chat Completions, Anthropic format, and Responses API. List prices from August 16 use peak and off-peak bands, with Flash peak output at $1.32 per million tokens. Independent index snapshots from Reuters and Artificial Analysis put Flash in a usable intelligence band. OpenRouter and OpenCode volume figures reported by Global Times show real production traffic after launch.

Our read: Over the next 12–36 months, small and mid-size U.S. shops will treat Flash-class open weights as the default worker for tool-using back office jobs, and they will keep a closed frontier model only for the tickets Flash fails. The constraint that broke in July — agent skill without a new architecture — will be copied. Prices will not stay at the May promotional valley; the August peak clock is the tell. Shops that already isolate model calls behind a workflow layer will absorb those copies and those price moves as config. Shops that baked a single vendor into every prompt will pay for a rewrite. If cache-hit rates stay high and off-peak windows stay wide, Flash remains the rational default even after another hike, because the 3.00× Pro step and the ~38× Fable 5 step still dwarf a 2× Flash bump. That last sentence is a forecast, not a measurement.

FAQ

What is DeepSeek V4-Flash?

DeepSeek V4-Flash is DeepSeek's cheaper open-weight mixture-of-experts model, officially released on July 31, 2026 as DeepSeek-V4-Flash-0731, built to use tools, run code, and finish multi-step agent jobs at a lower token price than V4-Pro. Call it as deepseek-v4-flash on the DeepSeek API.

Is DeepSeek V4-Flash open source?

The official 0731 weights and repository are published on Hugging Face under the MIT License, which allows use, copy, modify, and sale of the software with the copyright notice attached. "Open weights" is the accurate phrase; training data is not in that download.

How does DeepSeek V4-Flash differ from V4-Pro?

Flash activates 13 billion parameters per token against Pro's 49 billion, at 284 billion vs 1.6 trillion total, per Caixin and the paper. Official Flash beat the Pro preview on agent benches in July; official Pro-0813 later scored 87.9 vs Flash's 82.7 on Terminal Bench 2.1 in the changelog. Pro costs 3.00 times Flash on peak output at current list rates.

What did the July 31 update change if the architecture stayed the same?

It changed the trained behavior, not the skeleton. DeepSeek says 0731 was only re-post-trained, and the 0731 card shows Terminal Bench 2.1 moving from 61.8 (preview) to 82.7 (official). That is more practice on agent tasks, not a new model family.

Can a small business run DeepSeek V4-Flash locally?

Yes if you have the accelerators and ops time; otherwise use the API. The 0731 card publishes vLLM and SGLang serve commands, and the MIT license allows it. A two-truck HVAC shop should start on the API. A hospital IT team with a GPU cluster can host.

When did API prices change, and what is off-peak?

New V4 peak and off-peak rates take effect at 16:00 UTC on August 16, 2026, per DeepSeek's GA note. Off-peak is half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, on the pricing page.

Does Flash work inside existing coding agents?

Yes. The official Flash API natively supports the Responses API format and Codex setup, per the changelog and Responses API guide. Claude Code, GitHub Copilot, and OpenCode are named as compatible paths on the first-call page.

Glossary

• DeepSeek V4-Flash: DeepSeek's cheaper V4 model (284B total / 13B active); official API build is DeepSeek-V4-Flash-0731, released July 31, 2026.

• Post-training: Extra training after the base model is built; DeepSeek says this is the only change from Flash preview to 0731.

• Mixture-of-experts (MoE): A model that stores many specialist networks and turns on only a slice per token; that is why Flash can be large on disk and cheap per token.

• Context window: How much text the model can see at once. V4 Flash and Pro list 1 million tokens, with a 384K maximum output.

• Thinking effort: The low / high / max knob that sets how much hidden reasoning the model spends before it answers.

• Tool calling: The loop where the model names a function, your software runs it, and the model continues from the result.

• Peak / off-peak pricing: DeepSeek's weekday UTC price bands; off-peak list rates are half of peak.

• Responses API: An OpenAI-style request format DeepSeek now accepts so coding agents such as Codex can talk to Flash without a custom adapter.

What to do this week

If you already have an agent graph, change the model string to deepseek-v4-flash, set thinking effort to high, turn on tools, and log token counts for one week. Compare fail rates against your current model on the same 20 tickets, not on a chatbot arena. If you pay DeepSeek directly, shift batch jobs out of the published peak windows. If the files are sensitive, trial the 0731 checkpoint on a box you control before you send production data.

US Tech Automations customers who already isolate the model step can run that trial as a config change and keep the rest of the graph — retries, approvals, writes to CRM or ledger — untouched. When you want the worker behind the graph, not another chatbot, open the agentic workflows platform and plug Flash in as the cheap default.

For the company homepage, start at https://ustechautomations.com/. For the model itself, start at the official Flash 0731 card and the pricing page.