Using DeepSeek Models Through an OpenAI-Compatible API Gateway

·4 views

DeepSeek offers an OpenAI-style endpoint, but developers must decide where to store the base URL, how to manage model names, and what to test before routing production traffic. The guide uses the Tokuse gateway as an example, which lists deepseek-flash and deepseek-v4-pro and exposes an OpenAI-compatible base URL. Because existing OpenAI SDKs can usually be pointed at a new base URL and model name, applications can keep their current client structure while the gateway handles model resolution and provider credentials. Compatibility is not complete, however, so streaming, tool calls, JSON output, error codes, and token accounting should be tested with the exact SDK and model in use. The article also recommends keeping API keys in server-side environment variables, starting with the cheaper flash model for routine tasks, and judging model choice by successful task completion and cost per success.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Goldman Sachs Reportedly Pulling Claude Code From Its Bankers and Traders

Goldman Sachs is reportedly removing access to Anthropic's Claude Code tool for its bankers and traders. The move was reported by The State of AI, which the Hacker News post links to. The available material provides only the headline, link, and discussion metadata, so the specific reasons behind the decision are not confirmed here. Further details would need to come from the original report.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Cloudflare retires 24 Workers AI models on November 20, 2026

Cloudflare will stop serving 24 Workers AI models on November 20, 2026, covering 16 text-generation models plus eight models for image generation, speech, transcription, vision and embeddings. Requests to the retired models will return HTTP 410 Gone with error code 5028, and Cloudflare has not said that traffic will be automatically redirected to a newer model. For most text models the company suggests general-purpose replacements rather than one-to-one mappings, while the other eight models each have a named successor. Embedding users face the most difficulty, since the replacement embedding model was not yet available when the notice was posted, and vectors from different embedding models are generally not directly comparable. A smaller change also takes effect: Durable Objects will now retry failed calls by default.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Ahmedabad student's restaurant software models orders as shared state transitions

A Computer Engineering student in Ahmedabad is building RestaurantOS, a restaurant operations product from his company Nirvaan Tech. The system links the owner, waiter, kitchen, and cashier around a single order that moves through a defined sequence from placement to payment. The author says the main challenge was defining which role may change an order and when, rather than the interface itself. Instead of separate dashboards, each order is treated as a state machine in which each role can only advance it according to its responsibilities. The author is asking other developers in logistics, healthcare, and similar fields for feedback on where this approach may fail.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Kubernetes 1.35 in-place memory resize may not need a restart on ZGC and Shenandoah

A common recommendation is to set resizePolicy to RestartContainer for memory on every JVM container in Kubernetes 1.35. The author argues this is right for raising the heap ceiling, which is fixed by -Xmx at startup, but not for shrinking and regrowing within a ceiling chosen at launch. In their Linux container tests on ZGC and Shenandoah, a container's limit dropped from 900 MB or 700 MB to 300 MB without a restart, provided the JVM released memory first. Lowering the limit before the JVM released memory caused OOM kills every time. The author notes that the ZGC and Shenandoah SoftMaxHeapSize flag, which can be changed at runtime, makes this approach possible. The tests were run on Eclipse Temurin 21 with Docker Desktop using cgroup v2, which is a simulation rather than a full kubelet-driven pod resize.

0 comments Read more at DEV Community