DeepSeek V4.1 Flash Review: Benchmarks, Pricing, and What’s New

DeepSeek V4.1 Flash is DeepSeek’s September 10, 2026 release, adding native image input, a smaller cache footprint, stronger vendor-reported agent scores, and lower API rates. It best suits developers running high-volume or input-heavy agents. The main caveat: DeepSeek ran the published benchmarks, and V4 Pro still leads selected frontier reasoning tests.

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is a multimodal, open-weight model built for coding, tool use, visual understanding, and long-context agent workflows. DeepSeek released it on September 10, 2026, publishes the weights as DeepSeek-V4.1-Flash, and exposes the hosted model through the deepseek-flash API name.

The release matters because it changes more than benchmark scores. Flash can accept text and images, offers a one-million-token context window, and substantially reduces the memory required for its key-value cache. DeepSeek also cut its hosted API rates compared with the outgoing V4 Pro tariff. That combination targets a practical bottleneck: keeping long-running agents affordable when prompts, repository context, screenshots, and tool history accumulate.

Thinking mode is enabled by default, although non-thinking mode is available. JSON Output and Tool Calls are supported. DeepSeek also provides Responses API and Anthropic API compatibility, which reduces migration work for applications already structured around those interfaces.

What’s New in DeepSeek V4.1 Flash?

A 552B MoE backbone with asymmetric compute

Selon la DeepSeek V4.1 technical report, the model uses a 552-billion-parameter mixture-of-experts backbone with a 20-layer causal encoder and a 20-layer decoder. The full parameter count is not the amount used for every token: DeepSeek reports roughly 8B active parameters during prefill and 16B during decoding.

That asymmetric design gives prefill and generation different compute budgets. Prefill is the stage that absorbs the prompt, documents, images, and prior conversation; decoding produces the answer token by token. Lower active compute during prefill is particularly relevant to repository-scale and research-agent workloads that spend heavily on reading before they write.

A much smaller KV cache

The report lists a global KV cache size of 890 bytes per token, about one quarter of V4 Flash’s cache in high-bandwidth memory. DeepSeek says the persistent-cache requirement is about one eighth of the previous generation. For a one-million-token context, this is a deployment feature rather than an abstract architecture detail: cache size affects how many long sessions a server can hold and how much memory each one consumes.

Native image input and multimodal pretraining

V4.1 Flash accepts images natively instead of using a separate vision-branded endpoint. DeepSeek says the model was pretrained on a 45-trillion-token multimodal corpus. In practice, that makes screenshots, charts, scanned tables, interface states, and visual debugging inputs part of the same agent conversation as code and text.

DeepSeek V4.1 Benchmarks: What the Official Results Show

DeepSeek’s strongest V4.1 Flash claims are concentrated in coding and agent benchmarks. The vendor evaluated the instruct model with experimental effort 100, exposed publicly as maximum reasoning effort, using temperature=1.0 et top_p=0.95. Code-agent tests used named harnesses and a one-million-token context; visual-agent tests used Claude Code with a 512K-token context.

The pattern is useful but narrower than a blanket “better model” claim. V4.1 Flash leads V4 Pro on the five coding, terminal, security, and automation rows shown above. V4 Pro remains ahead on GPQA Diamond and the text-only subset of Humanity’s Last Exam. Developers comparing a wider coding shortlist can use our Guide sur le meilleur modèle d'IA pour la programmation, but application-level evaluation still matters more than one table.

Reasoning effort also changes cost and outcome. In DeepSeek’s report, raising experimental effort from 25 to 100 increased DeepSWE v1.1 from 66.0 to 74.2 and Terminal-Bench 2.1 from 82.4 to 90.6, while consuming roughly 2.5 times as many output tokens. The report’s eight agent configurations also produced materially different results from the same checkpoint, so framework choice is part of the measured system.

DeepSeek says extreme-input robustness is not fully characterized. The report notes that CSA2 selection errors and approximate sliding-window-attention state reconstruction may reduce capability in boundary cases outside the test suite. It also acknowledges a remaining gap on the hardest frontier reasoning and edge-case tasks. Those qualifications should travel with the headline scores.

Two reproducible tasks for evaluating V4.1 Flash

The following tasks test concrete behaviors without presenting an unrun prompt as evidence. Use the exact prompt, keep the model and effort setting in your record, and score the answer against the listed criteria. Results from one run should be reported as a dated observation, not a general benchmark.

Read the attached DeepSeek technical-report benchmark table. Return JSON only, with this exact top-level shape:
{
  "Codeforces": {"DeepSeek V4.1 Flash": ..., "DeepSeek V4 Pro": ..., "DeepSeek V4 Flash": ..., "leader": "..."},
  "Terminal-Bench 2.1": {"DeepSeek V4.1 Flash": ..., "DeepSeek V4 Pro": ..., "DeepSeek V4 Flash": ..., "leader": "..."},
  "DeepSWE v1.1": {"DeepSeek V4.1 Flash": ..., "DeepSeek V4 Pro": ..., "DeepSeek V4 Flash": ..., "leader": "..."},
  "GPQA Diamond": {"DeepSeek V4.1 Flash": ..., "DeepSeek V4 Pro": ..., "DeepSeek V4 Flash": ..., "leader": "..."}
}

Extract these four rows only. Preserve every number exactly as printed. Do not infer unreadable or missing values; use null instead. Choose `leader` by the highest numeric value in each row.

You are reviewing a TypeScript concurrency helper. It is intended to run at most `limit` tasks at once, preserve result order, and reject if any task rejects.

```ts
export async function mapLimit<T, R>(
  items: T[],
  limit: number,
  fn: (item: T, index: number) => Promise<R>,
): Promise<R[]> {
  const results = new Array<R>(items.length);
  let next = 0;

  async function worker() {
    while (next < items.length) {
      const index = next;
      await Promise.resolve();
      next++;
      results[index] = await fn(items[index], index);
    }
  }

  await Promise.all(
    Array.from({ length: Math.min(limit, items.length) }, worker),
  );
  return results;
}
```

Return exactly three sections:
1. Root cause
2. Minimal unified diff
3. One Vitest regression test that measures peak concurrency and result order

Do not redesign the API. Also reject `limit <= 0` with `RangeError`.

DeepSeek V4.1 Flash Pricing

DeepSeek API pricing uses two schedules, and all figures below are in US dollars per one million tokens. Peak periods are 01:00-04:00 UTC and 06:00-10:00 UTC, Monday through Friday. Every other time is off-peak. Cached input, uncached input, and output are billed at different rates.

For a simple arithmetic example, an off-peak request with 800,000 uncached input tokens and 20,000 output tokens would cost about $0.132: $0.12 for input plus $0.012 for output. The same token counts at peak rates would cost about $0.264. This calculation is illustrative, not a measured workload bill; cache hits, retries, tool turns, and reasoning length change the total.

Cached input is fifty times cheaper than uncached input in both schedules, making stable prompt prefixes and reusable context economically significant. The public DeepSeek API tariff is separate from GlobalGPT access and billing. Do not interpret these token prices as GlobalGPT subscription or credit prices.

DeepSeek V4.1 API, Vision, and Limits

The direct DeepSeek API uses deepseek-flash. Developers can call the standard endpoint at https://api.deepseek.com, the Responses API at the same base, or the Anthropic-compatible base at https://api.deepseek.com/anthropic. Function tools are supported in Responses. DeepSeek also documents the appliquer_le_patch custom tool; other Responses built-in tool types are ignored rather than executed.

Vision formats and placement rules

DeepSeek documents JPEG, PNG, GIF, and WebP image inputs. An inline image can be up to 32 MiB, while a file image can be up to 64 MiB. The documented shared limit allows up to 600 images in one request. That ceiling is not a recommendation: large batches increase prompt size, latency, and the difficulty of tracing which image supported an answer.

For the Responses API, image content belongs in user messages. Images placed in system ou assistant messages return HTTP 400. The compatibility layer does not accept general file inputs as input items, so applications should follow the documented image-input shapes instead of assuming complete OpenAI Responses feature parity.

Context, output, and concurrency limits

The maximum context is one million tokens, and the maximum output is 384,000 tokens. Input and output together must fit the context window. Flash allows 2,500 concurrent connections per account; requests above that account limit receive HTTP 429. Capacity planning should therefore track active connection occupancy and retry behavior.

DeepSeek V4.1 Flash vs V4 Pro

V4.1 Flash is the more practical default for high-volume coding agents, visual inputs, and long-context applications. V4 Pro remains relevant as a reasoning baseline during the short transition window, but it does not support vision and has lower documented concurrency. Our earlier Test du DeepSeek V4 Pro covers that checkpoint on its own terms; its hands-on results should not be relabeled as V4.1 evidence.

At off-peak rates, Flash’s uncached input is about 77% cheaper than the listed Pro rate, and Flash output is about 70% cheaper. Those percentages are arithmetic from the official tariff, not measured savings. Actual savings depend on cache reuse, output length, retries, and whether the workload needs Pro’s stronger results on the selected reasoning rows.

The more urgent distinction is operational. DeepSeek is keeping the deepseek-v4-pro name while temporarily changing the model behind it. A stable API name does not guarantee stable output behavior. Preserve representative V4 Pro responses, tool traces, latency, and failure cases before September 14, then run the same acceptance suite after routing changes. For the older checkpoint comparison framework, see DeepSeek V4 Pro vs Flash.

What Developers Are Saying

Early reaction is positive about throughput, cache efficiency, and agent scores, but the evidence is uneven. Some commentary concerns a preview endpoint rather than the final production checkpoint, and some reports summarize demonstrations instead of controlled tests. The most useful public discussion is therefore about where to test, not whether one launch-day number settles the decision.

These reactions point to three checks worth running on any real deployment: compare output quality at a fixed effort setting, watch whether long reasoning inflates output cost, and rerun regression cases when an alias changes the model behind an endpoint. Preview throughput and security results can guide test design, but they should not be combined with DeepSeek’s Table 3 as if they came from one controlled benchmark.

How to Access DeepSeek V4.1 Flash

There are three distinct access routes. Choose based on whether you need direct API control, self-hosted weights, or a browser workspace that makes model switching easier. Terms, pricing, and operational responsibility differ by route.

The direct API is the clearest choice when you need programmatic controls, usage accounting, and access to DeepSeek’s documented interfaces. Open weights are for teams able to validate inference support and operate a very large MoE model. GlobalGPT is the lower-friction route for interactive use and comparisons inside an all-in-one AI model workflow.

DeepSeek’s official launch post says the company will work with the open-source community on inference support and explore more deployment options. Treat that as a roadmap statement, not proof that every local serving stack already supports V4.1 Flash efficiently.

Who Should Use DeepSeek V4.1 Flash?

For broader model selection, the GPT-5.5 vs DeepSeek V4 comparison puts the one-million-token context and price position beside another frontier option. Teams focused on code-output style can also consult the DeepSeek vs ChatGPT Python code test. Those articles answer adjacent questions; neither substitutes for testing the exact V4.1 Flash endpoint used in production.

General users deciding between ecosystems may find the broader Comparaison entre DeepSeek et ChatGPT more useful than an architecture-level review. V4.1 Flash is most compelling when its specific advantages – long input, image understanding, tools, and low token cost – map to an actual workflow.

Verdict

Start with a small acceptance suite: one representative coding task, one image-grounded extraction task, one long-context case, and one tool workflow. Fix the reasoning setting, record model routing, and compare failure modes as well as average quality. That produces a decision you can defend without treating vendor benchmarks or launch-day reactions as guaranteed application performance.

Questions fréquemment posées

What is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is a multimodal, open-weight mixture-of-experts model released on September 10, 2026. It supports text and image input, text output, a one-million-token context window, thinking and non-thinking modes, JSON Output, and Tool Calls.

How much does the DeepSeek V4.1 Flash API cost?

Off-peak DeepSeek API rates are $0.003 per 1M cached input tokens, $0.15 per 1M uncached input tokens, and $0.60 per 1M output tokens. Peak rates are $0.006, $0.30, and $1.20 respectively. These are DeepSeek API prices, not GlobalGPT prices.

What API model name should I use for DeepSeek V4.1 Flash?

Utilisation deepseek-flash with the official DeepSeek API. DeepSeek also supports Responses API calls at https://api.deepseek.com and Anthropic-compatible calls at https://api.deepseek.com/anthropic.

What are the context, output, and concurrency limits?

DeepSeek V4.1 Flash has a one-million-token context window and a maximum output of 384,000 tokens; input plus output must fit inside the context. DeepSeek documents 2,500 concurrent connections per account, with HTTP 429 returned above the limit.

Is DeepSeek V4.1 Flash better than V4 Pro?

It depends on the task. In DeepSeek’s vendor-run table, V4.1 Flash leads V4 Pro on the displayed coding, terminal, security, and automation benchmarks, while V4 Pro leads GPQA Diamond and HLE text-only. Flash also costs less, supports vision, and allows more concurrent connections.

Does DeepSeek V4.1 Flash support images?

Yes. It accepts JPEG, PNG, GIF, and WebP inputs and produces text. DeepSeek documents up to 32 MiB for inline images, 64 MiB for file images, and 600 images per request under shared limits. Responses API images must be placed in user messages.

Are DeepSeek V4.1 Flash weights available to download?

Yes. DeepSeek publishes the DeepSeek-V4.1-Flash model weights on Hugging Face under the MIT license. The permissive license does not remove the infrastructure requirement: teams still need compatible serving software and enough hardware for the 552B MoE backbone.

Can I use DeepSeek V4.1 Flash on GlobalGPT?

Yes. GlobalGPT has a dedicated DeepSeek V4.1 Flash entry for browser-based use. Its access and billing are separate from DeepSeek’s direct API tariff, and platform availability does not independently verify DeepSeek’s published benchmark results.