Model

DeepSeek-V4.1 (Think Max)

OPEN

DeepSeek · DeepSeek-V4 family · released Sep 22, 2026

DeepSeek's September 22, 2026 incremental refresh of the V4-Pro line — same 1600B/49B MoE skeleton and 1M context, retrained post-training stack, filed as the three-config V4.1 family (Non-Think default / Think High / Think Max best). API price moves from $0.435/$0.87 to $0.50/$1.00 per Mtok — the first price increase in the V line, still far below every closed frontier model. The launch table's honest read: +1.5 to +3 points over V4-Pro across the shared rows, nothing that changes the open-vs-closed frontier picture by itself. Think Max additionally carries an independent ARC Prize ARC-AGI-2 run (49.3) — unusually quick third-party coverage for a DeepSeek release.

ReasoningCodingVisionFunction callingTool useAgentic

2568.8

Elo · rank #29

Parameters

1600B

Active params

49B (MoE)

Context

1M tokens

Architecture

1600B-total/49B-active sparse MoE (same skeleton as DeepSeek-V4-Pro), 1M-token context; text-only

License

MIT

Languages

—

API price (in/out)

$0.5 / $1

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-AGI-2Reasoning49.3%#19

best: GPT-6 Astra · 94.6%

BrowseCompAgents86.1%#9

best: Kimi K3.1 · 92.4%

CodeforcesCoding3287#1

best: this model · 3287

GDPval-AAAgents1612#15

best: Claude Fable 5 · 1932

GPQA DiamondReasoning91.3%#27

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning40.6%#33

best: Claude Opus 5 · 64.7%

LiveCodeBenchCoding94.2%#1

best: this model · 94.2%

MMLU-ProKnowledge88.4%#8

best: Claude Fable 5 · 91.5%

SWE-bench ProCoding57.9%#30

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding81.6%#12

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding71.8%#27

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA1088.0 GB8× H200 141GB

LoRA3408.0 GBbeyond 8× B200

Full fine-tune25680.0 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $25.34 (8× H200 141GB)

Plan a fine-tune →

DeepSeek-V4 family

Elo progression across releases

DeepSeek-V4-Flash (Non-Think)Apr 20261559.0 DeepSeek-V4-Flash (Think High)Apr 20262152.2 DeepSeek-V4-Flash (Think Max)Apr 20262233.7 DeepSeek-V4-Pro (Non-Think)Apr 20261606.7 DeepSeek-V4-Pro (Think High)Apr 20262303.8 DeepSeek-V4-Pro (Think Max)Apr 20262435.7 DeepSeek-V4-Flash-0731Jul 20262693.8 DeepSeek-V4.1 (Non-Think)Sep 20261735.3 DeepSeek-V4.1 (Think High)Sep 20262475.1 DeepSeek-V4.1 (Think Max)Sep 20262568.8

Resources

HFdeepseek-ai/DeepSeek-V4.1 DOCShttps://api-docs.deepseek.com/quick_start/pricing/APIAvailable via DeepSeek API

API price $0.5/$1 · each benchmark row carries its own source badge (see methodology)