anoman

DeepSeek

DeepSeek R1

modeldeepseek-r1
MidStabilStreamingReasoningBatchLong contextChina

Konteks

128K

Input / 1M token

$0.5500

Output / 1M token

$2.19

Rumus weighted token

1× provider × 4× tier

Tentang model ini

DeepSeek R1

DeepSeek R1 is the reasoning-specialist sibling of V3 — trained with reinforcement learning to think step-by-step before answering. It matches or beats OpenAI o1 on math and competitive programming while costing ~80% less.

Rilis

2025-01

Batas data latih

2024-07

Parameter

MoE 671B (37B active)

Paling cocok untuk

Kasus penggunaan

  • Competition-grade mathematics (AIME, MATH)
  • Algorithmic coding and competitive programming
  • Multi-step logical inference
  • Scientific reasoning + research summarization
  • Any workload where you'd otherwise reach for o1

Kelebihan

Yang dikerjakan dengan baik

  • AIME 2024: 79.8% (matches o1)
  • MATH-500: 97.3% — near-saturating the benchmark
  • Codeforces Elo: 2029 (96.3 percentile)
  • Emits explicit chain-of-thought you can inspect or hide
  • Open weights (MIT license) for self-host fallback

Keterbatasan

Pahami trade-off-nya

  • !Output token count is high — chain-of-thought adds 2–5× the tokens vs a normal chat model
  • !Slower TTFT — minimum ~3 s latency on first token while it thinks
  • !Less suited to conversational UX — pair with V3 for chat, R1 for hard problems

Benchmark

Skor terpublikasi

Skor dari model card resmi dan leaderboard publik. Makin tinggi makin baik kecuali disebutkan lain.

BenchmarkSkorMengukur
AIME 202479.8Competition math, high-school olympiad
MATH-50097.3Mathematics, competition + textbook
HumanEval96.3Python code generation, pass@1
MMLU90.8General knowledge across 57 subjects
Codeforces96.3Percentile rank vs human competitive programmers

Contoh kode

Chat completion dengan DeepSeek R1

from openai import OpenAI

client = OpenAI(
    base_url="https://api.anoman.io/v1",
    api_key="anm-sk-..."
)

response = client.chat.completions.create(
    model="deepseek-r1",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

Weighted token

weighted_tokens = raw_tokens × 1 (provider) × 4 (tier)

Paket Pro: 20 jt weighted token/bulan. Multiplier gabungan 4×: 5,000,000 raw token tersedia.

Alternatif

Model serupa di katalog kami

Beban kerja yang cocok dengan DeepSeek R1 sering cocok dengan ini juga. Bandingkan benchmark dan harga sebelum menetapkan default.

Gunakan DeepSeek R1 lewat Anoman.