DeepSeek

DeepSeek’s frontier models — V4 Pro and V4 Flash for the latest tier, R1 for reasoning, V3 for stable baseline.

Available models

Model ID	Parameters	Context	Speed	Best For
deepseek-v4-pro	1.6T (49B active, MoE)	1M	Medium	Top reasoning, complex coding
deepseek-v4-flash	284B (13B active, MoE)	1M	Fast	General + coding, best value V4
deepseek-r1	671B (MoE)	64K	Slow	Reasoning, math, analysis
deepseek-v3	671B (MoE)	160K	Medium	Previous-gen flagship, stable

V4 vs V3 — When to upgrade?

	DeepSeek V4 Pro	DeepSeek V4 Flash	DeepSeek V3
Tier	Flagship	Value	Previous-gen
Context	1M	1M	160K
Max output	65K	65K	8K
Speed	Medium	Fast	Medium
Reasoning	Native	Native	No
License	MIT	MIT	MIT
Release	Preview	Preview	Stable

Rule of thumb: Default to deepseek-v4-flash for new work — it gives you 1M context, native reasoning, and the same family behavior at a fraction of the cost. Upgrade to deepseek-v4-pro when quality matters more than price. Stay on deepseek-v3 if you have a stable production workload that’s already tuned to it.

V4 vs R1 — When to use which?

V4 (both Pro and Flash) has reasoning built in, so it’s broadly capable on the same tasks R1 was made for. R1 still has its place for pure step-by-step chain-of-thought work where you want the deeper, slower reasoning trace. Most users will find V4 covers both general and reasoning needs.

Recommendation

Default to deepseek-v4-flash — the new V4 baseline. 1M context, MIT license, native reasoning, lowest V4-tier price.

Step up to deepseek-v4-pro for top-tier reasoning, complex coding, or research-grade work.

Use deepseek-r1 for pure reasoning tasks where you want a slower, deeper chain-of-thought trace. It maps to the reasoning model alias.

from openai import OpenAI

client = OpenAI(base_url="https://kymaapi.com/v1", api_key="ky-...")

# General + coding — use V4 Flash
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Refactor this module and explain the tradeoffs."}]
)

# Top reasoning — use V4 Pro
response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Walk through the proof, then propose a tighter bound."}]
)

Model aliases

Alias	Resolves to
reasoning	deepseek-r1

# Using alias
model="reasoning"  # → deepseek-r1