anoman

DeepSeek

DeepSeek V3

modeldeepseek-v3
BudgetStableStreamingToolsBatchLong contextChina

Context

128K

Input / 1M tokens

$0.1400

Output / 1M tokens

$0.2800

Weighted tokens formula

1× provider × 1× tier

About this model

DeepSeek V3

DeepSeek V3 is a 671B-parameter Mixture-of-Experts model that activates 37B parameters per token. Designed for general-purpose chat, document analysis, and structured output, it punches well above its cost class on coding, reasoning, and long-context comprehension. The MoE architecture means inference is roughly as fast as a dense 37B model while quality matches dense 70B+ flagships.

Released

2024-12

Training cutoff

2024-07

Parameters

MoE 671B (37B active)

Best for

Use cases

  • Cost-sensitive production chat
  • Long-context document analysis (128K window)
  • Code generation + review
  • Structured-output extraction (JSON, function calls)
  • Multilingual workloads (Chinese, English, code)

Strengths

What it does well

  • Top-tier coding ability for the price (HumanEval 89%)
  • Strong on Chinese-language benchmarks (CMMLU 88.8)
  • Native function calling + JSON mode
  • 128K input + 8K output — fits most agent loops
  • Mixture-of-Experts keeps inference cheap at scale

Limitations

Know the trade-offs

  • !Knowledge cutoff is mid-2024 — agents may need search augmentation for fresh facts
  • !Reasoning depth trails specialist reasoners (use DeepSeek R1 for math-heavy work)
  • !Outputs longer responses than GPT-4o by default — set max_tokens explicitly for cost control

Benchmarks

Published scores

Scores from official model cards and public leaderboards. Higher is better unless noted.

BenchmarkScoreMeasures
MMLU87.1General knowledge across 57 subjects
MMLU-Pro75.9Harder MMLU variant, 10 options per question
HumanEval89.0Python code generation, pass@1
MATH-50089.3Competition mathematics
DROP90.1Reading comprehension + numeric reasoning

Code example

Chat completion with DeepSeek V3

from openai import OpenAI

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

response = client.chat.completions.create(
    model="deepseek-v3",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

Weighted tokens

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

Pro plan: 20M weighted tokens/month. Combined multiplier 1×: 20,000,000 raw tokens available.

Alternatives

Similar models in our catalog

Workloads that fit DeepSeek V3 often fit these too. Compare benchmarks and pricing before locking in a default.

Use DeepSeek V3 through Anoman.