DeepSeek-V4.1-Flash review, pricing and limits

Best suited for cost-sensitive agentic coding and long-document workloads that need a huge context window without frontier-model pricing. It scores 90.6 on Terminal-Bench 2.1 and outputs at roughly 218.7 tokens per second, replacing DeepSeek's own V4-Pro as the vendor's recommended default in September 2026.

DeepSeek-V4.1-Flash is a multimodal AI model that scores 40 on the Artificial Analysis Intelligence Index, roughly double the median for similarly sized open-weight models. Built on a new Causal Encoder-Decoder Mixture-of-Experts architecture, it natively processes text and images across a 1M-token context window, positioned by DeepSeek as the replacement for its V4-Pro flagship.

Provider: DeepSeek · Family: DeepSeek V4

More about DeepSeek on HokAI

Context window: 1,048,576 tokens · Max output: 384,000

Input modalities: text, image, tool-calls · Output: text, tool-calls

About DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash was built by DeepSeek, the Hangzhou-based AI lab founded by Liang Wenfeng in 2023, and released in September 2026. It replaces DeepSeek-V4-Flash-0731 (July 2026) as the smallest model in a new architecture family: a Causal Encoder-Decoder (CED) design that splits inference into a causal encoder stage and a decoder stage, on a 552-billion-parameter Mixture-of-Experts backbone with 1 shared expert plus 384 routed experts per layer, 6 activated per token. The asymmetric design activates 8B parameters on input and 16B on output, cutting KV-cache memory sharply versus its predecessor. DeepSeek positions it to replace V4-Pro entirely, saying testing put it ahead on performance, cost, speed and completion time; V4-Pro is being phased out. On the vendor's instruct, max-reasoning-effort figures, V4.1-Flash scores 90.9 on GPQA Diamond and 90.6 on Terminal-Bench 2.1, up from the July predecessor's 82.7. DeepSeek's own DeepSWE v1.1 coding eval rose from 54.4 to 74.2, with a 3,471 Codeforces rating and 63.9 on Humanity's Last Exam (with tools). Artificial Analysis puts its Intelligence Index at 40, roughly double the 18 median for similarly sized open-weight models, with output speed of 218.7 tokens/second and a blended price of $0.18 per 1M tokens. The base, pre-instruct checkpoint separately reports 74.1 MMLU-Pro and 79.4 HumanEval. The model ships a 1,048,576-token context window (max output 384,000 tokens); no independent long-context recall benchmark had been published for this release at launch. Input is text and images: a purpose-built DeepSeek-ViT vision encoder converts images into embeddings processed jointly with text from the start of pretraining. Output is text only, with no native audio or video. The API is OpenAI-compatible with function calling (tools/tool_choice) and a JSON response_format mode with no schema enforcement, requiring the word JSON in the prompt. Reasoning uses a continuous reasoning_effort value from 1 to 100, not a fixed enum; low/high/max aliases map to 50/75/100, and thinking defaults on at effort 50 if unset. DeepSeek prices V4.1-Flash (API name deepseek-flash) on a peak/off-peak schedule, with rates doubling during two short UTC windows on weekdays; see pricing for exact per-token rates. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to V4.1-Flash at the same rate. Also available via Azure AI Foundry, Vertex AI, AWS Bedrock, Fireworks AI and OpenRouter. Weights are on Hugging Face under the MIT license; the native FP4+FP8 checkpoint needs an estimated 170-175GB of VRAM, versus roughly 90-100GB with community INT4 quantization at some cost to accuracy. DeepSeek has not published a system card or training-data cutoff date for V4.1-Flash. Independent testing of the DeepSeek line finds it answers most topics freely but refuses or deflects questions on China-sensitive subjects, including Tiananmen Square, Taiwan's political status and Xinjiang; one tracker found paired prompts scoring as more censored rose from 79% to 88% between the July preview and production release. DeepSeek's Open Platform Terms of Service state no specific training or retention policy for API data; its consumer-app privacy policy allows an opt-out and says personal data is processed and stored in mainland China. The mix of a 1M-token context window, sub-$0.20 blended pricing and a sharp coding/agentic benchmark jump suits cost-sensitive agentic coding, long-document processing and high-volume workloads where MIT-licensed weights matter for self-hosting or fine-tuning. A missing audio/video modality, no independently verified frontier benchmark suite, and no formal compliance documentation are the main reasons a regulated or voice-first team would pick GPT-5 or Claude-class models instead.

Pricing

Off-peak (all hours outside the windows below): $0.15 per 1M input tokens (cache miss), $0.003 per 1M (cache hit), $0.60 per 1M output. Peak hours (01:00-04:00 and 06:00-10:00 UTC, Mon-Fri) double every rate to $0.30 / $0.006 / $1.20. No free API tier; DeepSeek's consumer chat app is free separately from the API.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Causal Encoder-Decoder Architecture: Runs a dedicated encoder stage ahead of the decoder stage, activating only a small fraction of the full Mixture-of-Experts backbone on each token instead of running every expert.
• 1,048,576-Token Context Window: Supports over 1 million tokens of context, with a documented API output cap of 384,000 tokens.
• Native Vision Encoder: A purpose-built DeepSeek-ViT vision encoder feeds image embeddings into the model jointly with text, rather than as a bolted-on adapter.
• Continuous Reasoning Effort Control: Accepts any integer reasoning_effort from 1 to 100, with API aliases low/high/max mapping to 50/75/100, letting callers trade latency for accuracy in fine steps.
• Peak/Off-Peak API Pricing: Bills less during most of the week and doubles the rate during two short daily UTC windows on weekdays; see pricing for the exact per-token rates.

Pros

• Posts a 3,471 Codeforces rating and sizable gains on DeepSeek's own coding and reasoning evals versus its July 2026 predecessor.
• MIT-licensed weights on Hugging Face allow free commercial self-hosting and fine-tuning, unusual among models with frontier-adjacent coding scores.
• A context window far larger than most competitors and off-peak API pricing among the cheapest available for a vision-capable model at this benchmark tier.

Cons

• No published SWE-bench Verified, AIME 2025, ARC-AGI-2 or LMArena score for this release, making direct frontier comparisons hard.
• Text and image input only; no native audio or video modality.
• No public system card or training-data cutoff date, and independent audits find it refuses or deflects a large share of China-sensitive political questions.

Benchmarks

• MMLU-Pro: 74.1% vendor-reported · 13 Sep 2026 — A harder version of the 57-subject knowledge exam, % correct.
• HumanEval: 79.4% vendor-reported · 13 Sep 2026 — Small programs that must pass hidden tests, % passing.
• GPQA Diamond: 90.9% vendor-reported · 13 Sep 2026 — PhD-level science questions that are hard to search for, % correct.
• Terminal-Bench 2.1: 90.6% vendor-reported · 13 Sep 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 63.9% vendor-reported · 13 Sep 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 40 cited: Artificial Analysis · 13 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $0.18/M cited: Artificial Analysis · 13 Sep 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 219 tok/s cited: Artificial Analysis · 13 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit DeepSeek-V4.1-Flash Official Page