DeepSeek V4 Flash on VM0. Agent-grade coding at the cheapest price tier
DeepSeek's open-weight 284B Mixture-of-Experts model with 13B active per token. The 0731 build beats V4 Pro on every agent benchmark DeepSeek published, at $0.14 / $0.28 per 1M tokens.
1M tokens · Text / Code · Prompt cache
DeepSeek V4 Flash is the efficiency half of DeepSeek's V4 generation: a 284B-parameter Mixture-of-Experts model that activates 13B parameters per token, released under the MIT License with open weights. The 0731 build that shipped on 31 July 2026 kept the architecture unchanged and re-ran post-training on agent data, which moved it past DeepSeek V4 Pro (Preview) on all nine agent benchmarks DeepSeek published — 82.7 against 72.1 on Terminal-Bench 2.1, and 54.4 against 12.8 on DeepSWE.
Vendor list price is $0.14 / $0.28 per 1M tokens, with cache hits at $0.0028 / 1M and no charge for cache writes, which makes it the cheapest reasoning model in VM0's current lineup. It carries a 1M-token context window, up to 384K of output, and thinking mode on by default. It is text-only: no vision. Reach for Claude Sonnet 4.6 when a step needs images or the Claude Code framework, and for GPT 5.6 Luna when you want the cheap tier of the OpenAI family instead.
What is DeepSeek V4 Flash?
July 31, 2026 (DeepSeek-V4-Flash-0731, public beta) · The efficiency half of the DeepSeek V4 family: 13B active parameters against V4 Pro's 49B, post-trained for agent work.
DeepSeek V4 Flash first appeared on 24 April 2026 as the smaller half of the V4 family — a 284B-parameter MoE with 13B active per token and a 1M-token context window, next to the 1.6T-parameter V4 Pro. It left preview on 31 July 2026 as DeepSeek-V4-Flash-0731, a public-beta build that kept the architecture untouched and re-ran only the post-training pass, this time on agent-flavoured data.
That re-training is the whole story of the release. On DeepSeek's own numbers the 0731 build beats V4 Pro (Preview) on all nine published agent benchmarks while activating roughly a quarter of the parameters: Terminal-Bench 2.1 goes from 61.8 in the Flash preview to 82.7, DeepSWE from 7.3 to 54.4, Toolathlon-Verified from 49.7 to 70.3. Several of those land within a few points of Claude Opus 4.8 — 82.7 against 85.0 on Terminal-Bench 2.1, 25.2 against 25.7 on Agents' Last Exam — at a fraction of the price.
The caveats are worth stating plainly. Every score is DeepSeek-reported and run with the DeepSeek Harness, which has not been released, so none of it has been reproduced independently, and two of the nine sets are DeepSeek's internal ones. The model is text-only, trails the frontier on graduate-level reasoning, and still lags V4 Pro on long-horizon multi-step tasks. Thinking mode is on by default, and thinking tokens bill as output.
What's notable about DeepSeek V4 Flash
Headline architecture and capability features.
V4 Flash is a sparse Mixture-of-Experts model: 284B total parameters with 13B activated per token, where each MoE layer holds 1 shared expert and 256 routed experts at an intermediate dimension of 2048, and 6 routed experts fire per token. The first three MoE layers use hash routing, and multi-token prediction depth is 1. The published 0731 checkpoint is 304B parameters because it ships a speculative-decoding draft module on top of the 284B base. It supports a 1M-token context window with up to 384K of output, thinking and non-thinking modes, and a reasoning_effort parameter with low, high and max levels. Weights are MIT-licensed and ungated.
Specs at a glance
DeepSeek V4 Flash benchmarks
DeepSeek-reported figures from the DeepSeek-V4-Flash-0731 model card, run with the DeepSeek Harness in minimal mode at max reasoning effort (temperature 1.0, top_p 0.95). The harness has not been released, so none of these has been reproduced independently, and DSBench-FullStack and DSBench-Hard are DeepSeek's internal test sets.
DeepSeek V4 Flash pricing
Provider list price, per 1M tokens.
How DeepSeek V4 Flash behaves in practice
Observed behaviour from production agent runs.
Agentic execution
The 0731 post-training targets exactly this: driving a terminal, calling tools in sequence, and finishing a task without a human in the loop. 82.7 on Terminal-Bench 2.1 and 70.3 on Toolathlon-Verified put it in frontier company for the price.
Cost profile
$0.14 / $0.28 per 1M tokens, cache hits at $0.0028 / 1M, and cache writes not billed at all. That is the cheapest position in the current lineup and roughly a third of V4 Pro's output price, so it is the model to point at high-volume work.
Long context
1M tokens in and up to 384K out, so repository-wide reading and long transcripts fit without chunking. Quality on long-horizon, many-step agent loops still trails V4 Pro.
Reasoning depth
Three reasoning_effort levels — low, high and max — trade latency for deliberation, and DeepSeek's published scores are all at max. Graduate-level reasoning is not its strength; that is where Claude Opus 5 and GPT 5.6 Sol stay ahead.
Modalities
Text and code only. Anything with a screenshot, a scanned PDF or a chart in the loop needs a vision model such as Claude Sonnet 4.6 or GPT 5.6 Luna.
Best agent tasks for DeepSeek V4 Flash
High-volume coding agents
Bulk PR review, test writing, lint-and-fix passes and scheduled refactors, where the same agent runs hundreds of times a day. At $0.28 per 1M output tokens the per-run cost stays low enough that volume stops being the constraint.
Terminal and tool-driven automation
The 0731 build's strongest published results are on terminal work and tool use, which is exactly what a build-fix, deployment-check or log-triage agent does all day.
Repository-wide reading
A 1M-token window takes an entire mid-size repository, a long incident timeline or a full set of design docs in one pass, so a migration or audit agent can reason over the whole thing instead of chunk by chunk.
The cheap layer under a frontier orchestrator
Let Claude Opus 5 or GPT 5.6 Sol plan and review, and hand the many mechanical steps — reading files, running commands, drafting patches — to V4 Flash. You pay the frontier rate only on the steps that decide the run.
When to skip DeepSeek V4 Flash
Skip V4 Flash on anything with images in the loop, since it has no vision, and on graduate-level reasoning, where the frontier models are still clearly ahead. Long-horizon runs that stay coherent over many hours remain V4 Pro's and Claude Opus's territory. And treat the published agent scores as vendor claims until the DeepSeek Harness ships: benchmark them on your own repository before you move a production agent across.
DeepSeek V4 Flash vs other models
DeepSeek V4 Flash vs DeepSeek V4 Pro
Same family, opposite trade. Pro is the 1.6T flagship with 49B active per token; Flash activates 13B and lists at a third of Pro's output price. On DeepSeek's published agent benchmarks the 0731 Flash build now scores higher than V4 Pro (Preview) on all nine, so the case for Pro is depth on long multi-step reasoning rather than agent throughput. V4 Pro is no longer offered on VM0 — its page here is reference material.
DeepSeek V4 Flash vs GPT 5.6 Luna
Both are the cheap tier of their family and both run on the Codex framework. Luna is multimodal and backed by OpenAI's ecosystem; Flash is text-only, open-weight, lists at under a quarter of Luna's output price, and posts far stronger published agent-benchmark scores. Pick Luna when you need vision or OpenAI-specific behaviour, Flash when the work is code and tools.
DeepSeek V4 Flash vs Claude Sonnet 4.6
Sonnet 4.6 is a core agent model with vision, the Claude Code framework and Anthropic's tool-routing reliability; Flash is a cost-saving, text-only model at roughly a fiftieth of Sonnet's output price. Keep Sonnet on the steps that decide a run and give Flash the volume underneath it.
Bottom line: should you use DeepSeek V4 Flash?
DeepSeek V4 Flash is the cheapest way to run a competent coding agent on VM0: frontier-adjacent agent benchmarks, a 1M-token window, and $0.28 per 1M output tokens. Verify the vendor's numbers on your own repository, keep vision and the hardest reasoning elsewhere, and it is hard to beat on cost per completed task.
Frequently asked questions
What is DeepSeek V4 Flash's context window?
1M tokens in, and up to 384K tokens of output per response. DeepSeek recommends the full 384K output ceiling at the high and max reasoning effort levels.
How much does DeepSeek V4 Flash cost?
Vendor list price is $0.14 per 1M input tokens and $0.28 per 1M output, with cache hits at $0.0028 per 1M and no charge for cache writes. On VM0 it sits in the $ price tier, the cheapest of the four. DeepSeek has announced a peak-hours policy that would double its own list prices between 09:00-12:00 and 14:00-18:00 Beijing time; it has not taken effect yet.
Is DeepSeek V4 Flash better than DeepSeek V4 Pro?
On the nine agent benchmarks DeepSeek published for the 0731 build, yes: it scores higher than V4 Pro (Preview) on every one, while activating 13B parameters against Pro's 49B. Pro keeps the edge on long-horizon multi-step reasoning. V4 Pro is no longer offered on VM0.
Does DeepSeek V4 Flash support vision?
No. It takes text and code only. Route screenshot-, chart- or PDF-driven steps to a multimodal model such as Claude Sonnet 4.6 or GPT 5.6 Luna.
What framework does DeepSeek V4 Flash use on VM0?
Codex. VM0 routes it through DeepSeek's Responses API, which V4 Flash is currently the only DeepSeek model to support. Run it as a Built-in model billed in VM0 credits, or bring your own DeepSeek API key.
Alternatives
Using DeepSeek V4 Flash on VM0
Two ways to access DeepSeek V4 Flash on VM0
VM0 supports DeepSeek V4 Flash as a Built-in model billed in VM0 credits, and through bring-your-own with a DeepSeek API key. The Built-in path uses VM0 Managed routing and the price tier explained below; the bring-your-own path bills you directly with the upstream vendor and skips the VM0 credit conversion entirely.
VM0's recommendation
VM0 positions DeepSeek V4 Flash as a cost-saving option rather than a core agent model. Use it to optimise unit cost on non-core work, such as bulk classification, pre-filters, latency-critical short replies, or pinned legacy agents, while keeping Claude Opus 4.7, Claude Opus 4.6, or Claude Sonnet 4.6 on the steps that decide the run.
Credits and the $ price tier
VM0 prices every Built-in model on a four-step credit scale — $, $$, $$$, $$$$ — shown as a badge in the model picker and on the price tier line of zero model ls. DeepSeek V4 Flash sits at $. That tier is what you spend from your VM0 credit balance; the vendor list price in the table above is what the upstream provider charges before VM0 converts it into credits.
Available on VM0 since July 31, 2026.