Back to blog

2026-09-02

How to call DeepSeek V4, GLM-5.3, and Kimi K3 with an OpenAI-compatible API from LatAm

If your stack already speaks Chat Completions, you do not need a separate SDK for every Chinese lab. DeepSeek, Z.ai (GLM), and Moonshot (Kimi) document OpenAI-compatible endpoints. NexoRouter exposes the same contract at https://api.nexorouter.com/v1: change base_url, the API key, and the published ID on the catalog.

This post summarizes facts from official announcements (July–August 2026). For prices, it only quotes strings that nexorouter.com/models showed on 2 September 2026. We do not paste long WeChat or vendor article text.

Diagram: LatAm app → NexoRouter → DeepSeek / GLM / Moonshot

What the vendors announced (official sources)

DeepSeek V4 and the API changelog

On 24 April 2026 DeepSeek published DeepSeek-V4 Preview: deepseek-v4-pro (1.6T / 49B active) and deepseek-v4-flash (284B / 13B active), 1M context by default, plus OpenAI Chat Completions and Anthropic interfaces. The API changelog on 13 August 2026 marks V4-Pro GA; 21 August adds experimental deepseek-v4-flash-vision-exp. Legacy IDs deepseek-chat / deepseek-reasoner were retired after 24 July 2026.

Z.ai GLM-5.3

On 14 August 2026 Z.ai published GLM-5.3: Frontier Coding with Emergent Cyber Capabilities. Same base as GLM-5.2; gains come from post-training. The API requires thinking enabled and reasoning_effort of low / high / max (disabled thinking is no longer supported).

Moonshot Kimi K3

Moonshot presents Kimi K3 on moonshot.ai (research dated 16 July 2026): 2.8T parameters, native multimodal, 1M context. The official Kimi K3 Quickstart sets model ID kimi-k3, reasoning_effort, and an OpenAI-compatible base_url. The model list states kimi-k2.5 and moonshot-v1 retired on 31 August 2026.

IDs and prices on the NexoRouter catalog (2 Sep 2026)

Copy the ID from Models, not from memory. Live strings that day:

| Catalog ID | Input / generation | Cached input | Output |

| --- | --- | --- | --- |

| kimi-k3 | $3.04 / 1M tokens | $0.30 / 1M tokens | $15.20 / 1M tokens |

| glm-5.3 | $1.22 / 1M tokens | $0.30 / 1M tokens | $4.26 / 1M tokens |

| glm-5.3-flash | $0.12 / 1M tokens | $0.03 / 1M tokens | $0.43 / 1M tokens |

| deepseek-v4-flash-0731 | $0.24 / 1M tokens | $0.05 / 1M tokens | $0.49 / 1M tokens |

| deepseek-v4-pro-sg | $2.73 / 1M tokens | $0.23 / 1M tokens | $5.47 / 1M tokens |

If an ID shows Pending, do not invent a rate: wait for a published string or smoke-test in Playground first.

Minimal call (OpenAI SDK)

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["NEXOROUTER_API_KEY"],
    base_url="https://api.nexorouter.com/v1",
)

r = client.chat.completions.create(
    model="deepseek-v4-flash-0731",  # or glm-5.3 / kimi-k3
    messages=[{"role": "user", "content": "In one sentence, what is an OpenAI-compatible gateway?"}],
)
print(r.choices[0].message.content)

Recommended steps:

1. Open Models, search the ID, copy the exact string.

2. Smoke-test in Playground with a low max_tokens.

3. Follow first API call and choose models.

4. Check Usage Logs before raising concurrency.

Verify the ID on /models and test in Playground

Migration footguns

  • DeepSeek: stop hard-coding deepseek-chat. Use the catalog ID (deepseek-v4-flash-0731, deepseek-v4-pro-sg, and so on). Official thinking effort is a request parameter, not a second legacy model name.
  • GLM-5.3: on Z.ai’s API, thinking.type: "disabled" fails. Fix that before swapping the model ID (see GLM-5.3 docs and the 14 Aug blog).
  • Kimi K3: thinking is always on; control effort with reasoning_effort. In multi-turn, return the full assistant message (including reasoning when present).

NexoRouter documents public Chat Completions. Do not copy Responses tutorials or other contracts that are not yet in /docs.

Why this matters in LatAm

Teams in Mexico, Colombia, Chile, Argentina, or Brazil often keep a single OpenAI client. A gateway with ES locale, visible USD list prices, and a filterable directory lowers the cost of switching when DeepSeek, Z.ai, or Moonshot publish a new ID. Flow: official announcement → check /models → Playground smoke → production.

Official sources cited

1. DeepSeek-V4 Preview — 24 Apr 2026

2. DeepSeek API Change Log — 13 and 21 Aug 2026 entries

3. GLM-5.3 (Z.ai) — 14 Aug 2026

4. Kimi K3 Quickstart (Moonshot)

5. Moonshot AI — Kimi K3, 16 Jul 2026

Prices: live capture of nexorouter.com/models on 2 Sep 2026. Re-check the directory before budgeting.