DeepSeek V4.1-Flash

V4.1-Flash is a sparse Mixture-of-Experts model with 552B backbone parameters and 40 Transformer layers, arranged as a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the encoder's final hidden states instead of from each decoder layer, which is what allows 8B active parameters during prefill and 16B during decode. Each MoE layer has 1 shared expert and 384 routed experts, with 6 routed experts active per token (Hugging Face config). Attention is DeepSeek's Compressed Sparse Attention 2 (CSA2): layers run in Full, Reindex or Reuse mode to share KV and sparse-attention indices, with a 128-token sliding window, a hierarchical sparse indexer in the decoder and FP4 KV caching. SWA Bounded Replay rebuilds short-window KV states by replaying recent tokens, so they never need to be stored on SSD. Other components are Single-Pass mHC residual mixing, an Engram conditional memory of 196B parameters accessed by token lookup (described separately from the 552B backbone), and DSpark speculative decoding. The vision encoder, DeepSeek-ViT, was trained from scratch and feeds a two-layer projector. Pre-training covered 45T multimodal tokens with sparse attention trained at 64K sequence length and context extended to 1M at the 34T-token mark. Post-training follows the standard SFT, RL and on-policy distillation recipe; DeepSeek says the changes are in data and environment synthesis, not algorithms. Vocabulary is 129,280 tokens. The published checkpoint is about 510 GB, with FP4 expert weights and FP8 elsewhere. Knowledge cutoff and training compute are not disclosed.

Coding (9.0) and agentic work (9.0) are the headline. On DeepSeek's max-effort table it scores 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, 54.8 on AutomationBench and 31.8 on Agents' Last Exam, all above V4-Pro. Independent Arena data points the same way: 12th on the agent leaderboard, ahead of Gemini 3.8 Flash (13th), and 1614 on Arena's separate Code arena, where V4-Flash scores 1580. The ceiling shows on science-heavy agent tasks: DeepSeek reports Terminal-Bench 4.0 at 31.2, against 51.8 for Opus 5. Reasoning (8.5) is strong but not top-tier: GPQA Diamond 90.9, but HLE 36.8, below V4-Pro's 42.7. Math (8.8) is close to the best open models, with MathArena Apex at 65.6, level with Kimi K3, and a Codeforces rating of 3471. Long context (8.0) keeps the 1M window, but DeepSeek publishes no instruct-model long-context score, and the base model trails V4-Pro-Base on LongBench-V2 (45.2 against 51.5). Multilingual (7.5): strong in Chinese (C-Eval 92.1 base), but base-model MGSM fell to 80.2 from V4-Flash-Base's 85.7. Vision (8.0) and document/OCR (8.0) are new: the base model scores 95.6 on DocVQA and 56.5 on MMMU-Pro, and with tools it reaches 78.9 on Chartography and 89.6 on BabyVision. DeepSeek itself says a gap remains to the largest closed models. Function calling (8.5) includes a strict JSON-schema mode in beta and holds up across agent harnesses. Instruction following (8.0) and creative writing (7.0) have no published scores; they follow the family's behaviour. Safety calibration (6.5) reflects PRC-law content handling and limited published safety work. There is no real-time data access through the API (0).

All rows except Arena are vendor-reported by DeepSeek, run at maximum reasoning effort (100) with temperature 1.0 and top-p 0.95. Code-agent results use DeepSeek's own harness in Minimal mode, except DeepSWE v1.1, which uses mini-SWE. Competitor figures in the vendor rows are DeepSeek's runs, not the other vendors' published numbers. The base rows measure the pre-trained model before post-training and are not comparable to instruct scores elsewhere in the catalog. NL2Repo-Bench is 65.4 in the changelog and technical report but 64.0 on the Hugging Face card. On Arena's agent leaderboard (as of 2026-09-21) V4.1-Flash ranks 12th (rank range 7–19). It is ahead of Gemini 3.8 Flash (13th), V4-Pro 0813 (16th), V4-Flash (22nd), GLM-5.3-Flash (25th), Qwen3.8-Flash-Next (26th) and GPT-5.6 Luna (27th). The Code arena is scored separately from LMArena's Text board and its Coding category, where V4.1-Flash was not listed on 2026-09-13, so the catalog's LMArena Elo columns stay empty for it. No Artificial Analysis Intelligence Index score is recorded for this model yet.

Independent throughput and time-to-first-token figures are not yet established, so both are left blank. By design it should be fast: 8B active parameters on input and 16B on output, DSpark speculative decoding, and a KV cache a quarter the size of V4-Flash's. DeepSeek claims faster inference, higher throughput and shorter total runtime than V4-Pro. In practice, latency depends on effort. Thinking is on by default at high effort, and DeepSeek's report shows that raising effort from 25 to 100 costs about 2.5x the output tokens. The 60–80 range recovers most of max-effort accuracy at under half the tokens. For interactive chat, low effort or non-thinking mode is the fast path. For long agent runs, budget for long traces at max effort. Latency tier: fast.

Headline prices on this page are DeepSeek's standard (peak) rates. Off-peak is exactly half and covers all other hours, weekends and Chinese public holidays in full. US business hours fall entirely inside off-peak, so a US-daytime workload in practice pays $0.15 in, $0.60 out and $0.003 for cache hits. The current prices took effect at 04:00 UTC on 2026-09-10. DeepSeek reserves the right to adjust them and has changed them twice since 2026-08-16.

The first-party API is at https://api.deepseek.com, with model name deepseek-flash, and speaks three formats: OpenAI Chat Completions, the Anthropic Messages API (/anthropic) and the OpenAI Responses API. DeepSeek adapted the Responses API for Codex. It runs on DeepSeek's infrastructure in China. The MIT open weights are on Hugging Face at about 510 GB, with FP4 experts and FP8 elsewhere. Plan on a single 8-GPU node with at least 80 GB per GPU; DeepSeek's reference inference code converts the checkpoint for eight-way tensor parallelism. There is no Jinja chat template: prompts are built with the bundled encoding.py or DeepSeek's deepseek-recipe Rust libraries. DeepSeek says it is working with the open-source community on inference-engine support and invites large deployments of around 2,000 GPUs to contact it. Neutral hosts on OpenRouter already serve it, including DeepInfra, Fireworks, Together, Novita, SiliconFlow, Baseten and Parasail, for teams that want managed access outside the PRC API. There is no first-party Bedrock, Vertex AI or Azure offering.

DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China and uses it to train and improve its models, and it lists a right to opt out of training use. The open-platform terms are governed by PRC law, and DeepSeek offers no zero-retention option or published compliance certifications (SOC 2, ISO 27001, HIPAA) for the API. The API applies content-safety handling per user_id, which also isolates KV cache and scheduling between a customer's end users. DeepSeek publishes no safety framework or refusal-rate data for this model. Its technical report calls the model's cybersecurity skill dual-use (CyberGym 88.1, an open-model best) and records exploit-seeking behaviour during testing, such as decompiling Ubuntu packages to find vulnerabilities. For regulated or sensitive data, the practical compliance path is to self-host the MIT weights or use a neutral host in the required region.

Access is through the OpenAI and Anthropic SDKs in Python and TypeScript, pointed at DeepSeek's base URLs, plus native OpenAI Responses API support tuned for Codex. DeepSeek names WorkBuddy (including CodeBuddy) and OpenCode as official launch partners, and its own DeepSeek Harness agent framework (Minimal, Standard and PTC modes) was used for most of its agent evaluations. For self-hosting, DeepSeek ships a reference inference implementation, a Python prompt encoder and the deepseek-recipe Rust libraries. Many independent providers serve it through OpenRouter, including DeepInfra, Fireworks, Together, Novita, SiliconFlow, Baseten and Parasail. Uptake is fast: the Hugging Face repository showed 512,120 downloads and 3,514 likes on 2026-09-21, eleven days after release.