Model
Model explorer

DeepSeek-V4.1-Flash-1017

OPEN
DeepSeek · DeepSeek-V4 family · released Oct 17, 2026

DeepSeek's October 17, 2026 checkpoint of V4.1-Flash: the same 552B-backbone causal encoder-decoder MoE (about 8B active in prefill, 16B in decode, native image input, 1M context) with further post-training, released as open weights under MIT. The deepseek-flash API alias, the web app and the legacy V4-Flash names now serve it at unchanged prices. DeepSeek reports gains over the September 10 checkpoint on every row it reran (HLE 42.6 vs 36.8, Terminal-Bench 4.0 42.3 vs 31.2, DeepSWE 77.0 vs 74.2, AutomationBench 60.5 vs 54.8) along with shorter outputs and runtimes. NOT FILED: Agents' Last Exam (35.4) is a pass rate, while the corpus column holds the score component; Chartography (83.0) is with tools, while the corpus column is without tools; ProgramBench, Apex Bench, BabyVision and ZeroBench have no corpus column; the text-only HLE figure (44.0) is a subset. GPQA, Codeforces, Terminal-Bench 2.1/3.0, CyberGym and SEC-Bench Pro were not rerun for this checkpoint.

ReasoningCodingVisionFunction callingTool useAgentic
2747.3
Elo · rank #19
Parameters
552B
Active params
16B (MoE)
Context
1M tokens
Architecture
552B sparse MoE, 40-layer causal encoder-decoder (1 shared + 384 routed experts, 6 routed active per token), ~8B active in prefill / 16B in decode, ~890 bytes/token global KV cache (about 1/4 of V4-Flash); natively multimodal (image+text in, text out); ~45T multimodal pre-training tokens
License
MIT
Languages
—
API price (in/out)
$0.3 / $1.2
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-AGI-1Reasoning94.5%#14
best: Claude Fable 5 · 98.5%
ARC-AGI-2Reasoning78.1%#15
best: GPT-6 Astra · 95.0%
AutomationBenchAgents60.5%#2
best: MiniMax M3.1 Flash · 62.8%
DeepSWECoding77.0%#4
best: Muse Wonder Full · 79.0%
best: Muse Spark 1.2 · 25.4%
Humanity's Last ExamReasoning42.6%#30
best: Kimi K3.1 Pro · 70.0%
NL2RepoCoding68.2%#1
best: this model · 68.2%
Terminal-Bench 4.0Agents31.8%#20
best: Claude Sonnet 5.5 · 70.6%
Vals Finance AgentAgents57.0%#15
best: Gemini 4 Argon · 65.4%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Permissive
QLoRA375.4 GB2× B200 192GB
LoRA1175.8 GB8× B200 192GB
Full fine-tune8859.6 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $5.72 (2× B200 192GB)
API price $0.3/$1.2 · each benchmark row carries its own source badge (see methodology)