DeepSeek V4.1-Flash is the September 10, 2026 release that makes the small model the headline. It is the smallest member of DeepSeek’s new architecture family — a 552B-parameter multimodal mixture-of-experts with a Causal Encoder-Decoder design that uses just 8B active parameters for input and 16B for output — yet DeepSeek reports it ahead of its own DeepSeek V4-Pro on performance, cost, speed and total runtime.
Weights are on Hugging Face under the MIT license, the API model is deepseek-flash, and the context window is 1M tokens with up to 384K output tokens.
Benchmarks
Vendor-reported at maximum reasoning effort (DeepSeek, September 10, 2026). The right-hand columns are DeepSeek’s own V4-Pro-0813 and V4-Flash; the other models are as shown in DeepSeek’s table.
| Benchmark | V4.1-Flash | V4-Pro-0813 | V4-Flash | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 | 74.0 | 73.0 |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 | 89.1 | 88.8 |
| Terminal-Bench 3.0 | 30.0 | 11.8 | 7.6 | 43.3 | 34.4 |
| Terminal-Bench 4.0 | 31.2 | 12.4 | 7.0 | 51.8 | 39.9 |
| CyberGym | 88.1 | 83.3 | 76.7 | — | 84.5 |
| AutomationBench | 54.8 | 43.2 | 37.7 | 50.3 | 45.8 |
| Agents’ Last Exam | 31.8 | 25.7 | 25.2 | 28.6 | 26.7 |
| GPQA Diamond | 90.9 | 92.4 | 89.9 | 93.4 | 94.1 |
| HLE (no tools) | 36.8 | 42.7 | 37.8 | 56.3 | 44.5 |
| HLE (with tools) | 63.9 | 60.0 | 51.5 | 63.6 | — |
The harness matters: DeepSeek reports 74.2 on DeepSWE v1.1 with the mini-SWE-agent harness and 72.6 with its own DeepSeek Harness in minimal mode, against 69.8 in Claude Code, 65.6 in Codex and 65.5 in OpenCode. Base-model knowledge results are mixed against the 1.6T V4-Pro-Base: MMLU-Pro 74.1 vs 73.5, but SimpleQA-Verified 42.3 vs 55.2.
Pricing
| Per 1M tokens | Peak | Off-peak |
|---|---|---|
| Input (cache miss) | $0.30 | $0.15 |
| Input (cache hit) | $0.006 | $0.003 |
| Output | $1.20 | $0.60 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour is off-peak. DeepSeek says prices were reduced with this release. The concurrency limit is 2,500.
Specs
| Spec | Value |
|---|---|
| API model | deepseek-flash (legacy deepseek-v4-flash names are routed here) |
| Architecture | 552B MoE, Causal Encoder-Decoder, CSA2 attention, Engram memory (196B parameters), DSpark speculative decoding |
| Active parameters | 8B for input, 16B for output |
| Pre-training | 45T multimodal tokens; context extended to 1M |
| Context window | 1M tokens; max output 384K |
| Modalities | Text and image in, text out |
| Reasoning | Thinking on by default; continuous effort setting 1–100 |
| License | MIT (weights and code) |
| API | OpenAI and Anthropic formats, Responses API, JSON output, tool calls, FIM (non-thinking) |
How it compares
- vs DeepSeek V4-Pro: the new Flash beats the 1.6T-parameter Pro on most agentic rows at a lower price; Pro stays ahead on HLE and GPQA.
- vs GLM-5.3 and Kimi K3: in DeepSeek’s table V4.1-Flash is ahead on DeepSWE v1.1, AutomationBench, CyberGym and Terminal-Bench 2.1 and 3.0; GLM-5.3 leads on Terminal-Bench 4.0 (37.9 vs 31.2), and V4.1-Flash trails both on HLE without tools.
- vs Mistral Large 4: DeepSeek’s DeepSWE score is higher and its API price lower, though harnesses differ.
- vs MiMo-V2.6-Pro: similar DeepSWE v1.1 (71.9 vs 74.2); MiMo is a 1.02T model with audio and video input.
- vs closed frontier models: it matches Claude Opus 5 and GPT-5.6 Sol on DeepSWE v1.1 but trails the newer Claude Opus 5.5 and GPT-6 generation on the hardest long tasks.
Limitations
- Smallest of the new family. DeepSeek has said V4.1-Pro is still to come.
- Vendor-reported. Scores come from DeepSeek’s harness at maximum effort.
- Time-based pricing. Peak-hour requests cost double.
- V4-Pro status has wobbled. DeepSeek first announced phasing out V4-Pro, then said it would continue serving it.
For the wider picture, see the 2026 LLM guide.
Best for
- Cost-sensitive coding and terminal agents (DeepSeek Harness, Claude Code, OpenCode, Codex)
- Long-context repository and document work at 1M tokens
- Vision-in-the-loop agents — charts, screenshots and documents
- Self-hosted or MIT-licensed deployments that need a near-frontier open model
Pros & cons
Strengths
- Open weights under the MIT license on Hugging Face, with a technical report and an inference reference
- Flagship-level agentic results from a small active footprint — DeepSWE v1.1 74.2%, Terminal-Bench 2.1 90.6%, CyberGym 88.1% — ahead of DeepSeek's own V4-Pro on most rows
- A new Causal Encoder-Decoder design with 8B active parameters for input and 16B for output, and a KV cache about one quarter the size (1/8 the SSD storage) of the previous generation, which cuts the cache-hit costs that dominate agent bills
- Native image understanding, a 1M-token context, up to 384K output tokens and a continuous 1–100 reasoning-effort setting
Limitations
- It is the smallest model of the new architecture — a V4.1-Pro is still to come, and DeepSeek has said it is phasing out V4-Pro
- Behind Claude Opus 5 on the hardest long tasks in DeepSeek's own table (Terminal-Bench 4.0 31.2% vs 51.8%) and on HLE (36.8% vs 56.3%)
- Peak/off-peak pricing makes cost depend on when you run — peak is double off-peak on weekday UTC mornings
- A 552B-parameter MoE plus a 196B-parameter Engram memory is still a multi-GPU deployment; DeepSeek points large deployments to 2,000-GPU-class clusters
Frequently asked questions
When was DeepSeek V4.1-Flash released?
DeepSeek released it on September 10, 2026 on the DeepSeek API as `deepseek-flash`, with weights and a technical report on Hugging Face under the MIT license.
How much does V4.1-Flash cost?
At peak, $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens; off-peak, half of that ($0.15 / $0.60). Cache-hit input is $0.006 peak and $0.003 off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.
What happened to DeepSeek V4-Flash and V4-Pro?
V4-Flash and V4-Flash-Vision-Exp are retired; their old model names are routed to V4.1-Flash. DeepSeek first said V4-Pro would also route to V4.1-Flash from September 14, then said it would keep serving V4-Pro after that date. The pricing page still lists V4-Pro-0813, and V4.1-Pro has not been released.
How big is V4.1-Flash, and how can it be so cheap?
It has 552B backbone parameters with an asymmetric design — 8B active for input processing and 16B for output — plus a KV cache a quarter the size of V4-Flash's. DeepSeek says that lets it serve more users at lower cost and pass the savings on.
Can it see images?
Yes. V4.1-Flash natively accepts images and text and generates text. DeepSeek reports Chartography 78.9%, BabyVision 89.6% and ZeroBench 49.0% with tools.