---
title: "Open-weight models over API: a quick guide"
description: "Seven open-weight models you can consume over API and when to use each: high-volume tool use, long context, coding, and clean licenses."
date: 2026-09-20
locale: en
tags: [ai]
summary: "A guide to seven open-weight models consumed over API and when to pick each: Qwen3.8-Flash-Next for cheap high-volume tool use, DeepSeek V4.1 Flash for cache and agentic loops, GLM-5.3-Flash multimodal, MiniMax M3 and Kimi K3 for million-token context, Kimi K2.7 Code for long-horizon coding, and Qwen3.8-Max for specs. All with published weights: switching providers reduces lock-in."
url: https://escribano.dev/en/blog/018-open-weight-models-api/
---
Seven open-weight models consumed over API: what each one does well and when to pick it. I verified prices on September 20, 2026; they expire fast (one already did while I was writing), so the table has no figures.

All seven have published weights: you can download them and, with the right infrastructure, run them yourself or move to another inference provider. Switching providers is not plug-and-play (URL, model ID, and parameters usually change), but it reduces an important dependency: if prices rise or an API goes away tomorrow, the weights remain available. It does not remove every dependency, but it changes the relationship quite a bit. It is the cognition layer I wrote about in the [post on digital sovereignty](/blog/005-digital-sovereignty/).

## The table

| Model | When to use it |
|---|---|
| Qwen3.8-Flash-Next | High volume with tools and multimodal input; the best price-performance on the list. This is the checkpoint served by the `qwen3.8-flash` API |
| DeepSeek V4.1 Flash | Tool-heavy, high-iteration agents (OpenClaw, Hermes, Moltis): among the best agentic scores on the list and by far the cheapest cache, which in those loops can dominate the bill. API ID: `deepseek-flash` (legacy aliases `deepseek-v4-flash` redirect). MIT license |
| GLM-5.3-Flash | Genuinely multimodal (image, video and files) under MIT |
| MiniMax M3 | Large refactors at low cost with tests or review behind: up to a million-token window (input plus output; MiniMax counts it that way) via its [MSA](https://github.com/MiniMax-AI/MSA) sparse attention architecture (long context stays usable), native multimodality, and adaptive reasoning that switches on only when needed |
| Kimi K2.7 Code | Long-horizon coding and systems engineering: completes tasks end to end with few reasoning tokens. Expensive, for when the task justifies it |
| Qwen3.8-Max | Writing the spec and breaking down the work (the cheap models do the implementing); also scores very well on Code Arena: WebDev. The API snapshot runs a step ahead of the downloadable weights (`Qwen/Qwen3.8-2.4T-A95B`) |
| Kimi K3 | Deep research and long-horizon tasks: cross-checks hundreds of sources with a million-token window (90.4 on BrowseComp in a 1M-context evaluation with no compaction, comparable to top-tier closed models). Always reasons, adjustable across three levels; expensive, for when the task justifies it |

## In practice

I've been using Cursor for several months and I'm happy with Composer 2.5, especially the speed in fast mode. Under the hood it's Kimi K2.5, and it has worked well for me both for programming and for homelab infrastructure (Docker, Podman, Quadlet and the like). That's why I'm now trying Kimi K2.7 Code in [OpenCode](https://opencode.ai/): I want to see if the Kimi family keeps that level outside the Cursor ecosystem.

I recently subscribed to OpenCode Go and plan to run Cursor, OpenCode and [OpenRouter](https://openrouter.ai/) in parallel for a while on side projects. OpenRouter, pay-as-you-go, is my plan B. At work I stick with closed models; in my own time I calibrate open ones in case they need to go to production someday, especially if inference prices keep rising (Anthropic is the clearest case). Open models can get more expensive too, but vendor lock-in is lower, and that's why this guide focuses on them.

The split I have in mind, still being tested: three workhorses for agents like OpenClaw and Moltis (I haven't tried Hermes yet): DeepSeek V4.1 Flash for cost and cache, GLM-5.3-Flash when the flow is multimodal or goes through Z.AI, and MiMo-V2.6-Flash when I want a million-token window in the same price tier. Day to day I still default to DeepSeek; cache and runtime detail belongs in the follow-up to this guide. Qwen3.8-Max for writing specs in [OpenSpec](https://github.com/Fission-AI/OpenSpec). I also want to validate GLM-5.3-Flash on frontend work: Z.AI pushes it hard on visual coding and UI reproduction. Kimi K2.7 Code is for backend (especially Go) and systems engineering: Quadlets, wiring up Tailscale or Zitadel, and similar tasks.

I save Kimi K3 for important code reviews: it always reasons and the price shows. With Qwen3.8-Max I learned the hard way: in a few hours I burned half my monthly OpenCode Go quota, largely because I let the context grow without watching it. Watching context matters for OpenSpec too, even though the work there is more bounded.

## What doesn't expire

Prices change. GLM-5.3-Flash cost half as much two weeks ago and the promotion ended on September 9 without notice; DeepSeek shipped V4.1 Flash on September 10 with new rates. Anyone reading this in December will find different numbers.