552B total, 8B active at prefill / 16B at decode (MoE: 1 shared + 384 routed experts, 6 routed active per token). First widely deployed release on the Causal Encoder-Decoder (CED) architecture - a 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder, replacing the decoder-only stack of V4. FP8 safetensors on the main branch (48 shards, 510GB).
-
Context: 1M native, with multimodal (text + image) input. CSA2 sparse attention; KV cache is FP4 at roughly 890 bytes per token, so a full 1M-context cache stays under 1GB - weights, not context, set the rig floor.
Announced 2026-09-09, live on the DeepSeek API as deepseek-flash with peak/off-peak pricing ($0.30/$1.20 peak per 1M in/out; off-peak half; cache-hit $0.006/$0.003). From 2026-09-14 04:00 UTC the deepseek-v4-pro alias routes to V4.1-Flash at Flash rates until V4.1-Pro launches. DeepSeek’s tech report puts the instruct model (max reasoning effort) at GPQA-Diamond 90.9, Codeforces 3471, Terminal-Bench 2.1 90.6, with Humanity’s Last Exam 36.8 and Terminal-Bench 4.0 31.2 - all vendor-reported pending independent evals.
The full comparison table. DeepSeek’s numbers put V4.1-Flash at the top of the agentic rows against the frontier set it cites: Terminal-Bench 2.1 90.6 (V4-Pro 0813: 87.9, GPT-5.6-Sol 88.8, Claude Opus 5 89.1), DeepSWE v1.1 74.2 (Sol: 73.0, Opus 5: 74.0), CyberGym 88.1, HLE with tools 63.9, and Automation-Bench 54.8. It gives up ground on GPQA-Diamond (90.9 vs Opus 5’s 93.4 and Sol’s 94.1) and NL2Repo-Bench (65.4 vs Opus 5’s 75.3). Every figure in the table is DeepSeek’s own reported number for every model shown - vendor-reported, pending independent replication.
Open weights are live on HuggingFace at deepseek-ai/DeepSeek-V4.1-Flash - FP8 safetensors on the main branch (48 shards, ~510GB), MIT license.
-
Local run: 4-bit quants land in the ~250GB range, so the target rig is 256GB+ unified memory (stacked DGX Sparks or a 512GB Mac-class box) - beyond a 128GB Spark. The model card ships an
inference/ folder for weight conversion; community quants exist in the llama.cpp, Ollama, and LM Studio ecosystems. First measured datapoint on the stacked-Spark shape: a community build (2026-09-11) runs V4.1 Flash across 3x DGX Spark 128GB (384GB unified, SGLang TP=3 EP=3) at 37.9 tok/s single-stream decode, with the full setup scripts published; a UI video posted the day before shows the same rig shape at ~28-37 tok/s in a thinking-mode chat.
1017 checkpoint (2026-10-17): a post-training refresh of this model, same architecture and same price. Weights live at deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT, fp8), OpenRouter serves it as the first provider route at the same $0.15/$0.60 per 1M off-peak. DeepSeek’s table shows the reasoning quality moving up across the agentic rows: Terminal-Bench 4.0 31.2 to 42.3, DeepSWE v1.1 74.2 to 77.0, ProgramBench 20.3 to 29.1, Automation-Bench 54.8 to 60.5, Agents’ Last Exam 31.8 to 35.4, Apex 72.1, NL2Repo 68.2, HLE 42.6 (44.0 text-only). The release claims shorter outputs and faster task completion, a performance-cost Pareto improvement; the Search and Fetch server-side tools are free until November 1, 2026 (per-request billing starts then). A user-measured datapoint on the new checkpoint: ~246 tok/s in an agentic harness session at 98% cache hit (2026-10-17).
Supersedes V4 Flash 0731 in the flash tier; V4-Pro routes to it from 2026-09-14. The deepseek-flash API name and legacy deepseek-v4-flash aliases now serve the 1017 checkpoint at Flash rates.