---
title: "Five open-weight LLM providers as Knights of the Zodiac"
description: "Saint Seiya in the 90s, Linux distros in the 2000s, and the open-weight catalogue: an analogy to keep Qwen, GLM, Kimi, DeepSeek and MiniMax straight."
date: 2026-10-18
locale: en
tags: [ai]
summary: "In the schoolyard, nobody agreed on which knight was the best, and open models are no different. A cheat sheet with the five Knights of the Zodiac protagonists to help you find your way."
url: https://escribano.dev/en/blog/021-llm-caballeros-zodiaco/
---
Once you decide to step away from frontier models and start digging into the open-weight catalogue, you find yourself in a proper jungle. Dozens of models, all publishing weights, none leading across the board. The question "which is the best open model?" has no clean answer.

The same thing happened to me with Linux distributions. When I got started in 2003, I tried several and even FreeBSD in search of the "perfect" one. Ubuntu did not exist yet (it arrived in 2004, and it is what we would call today the distro for beginners). Before that there was Debian, Red Hat, Mandrake, Slackware, Gentoo, SUSE: each with its community and its holy wars. Asking which one was best always ended in *flamewars*, because the answer depended on what you wanted it for.

And there is an even older feeling this reminds me of. TVE started airing the show in 1991, and in the schoolyard everyone had their favourite knight; mine was the Dragon. Nobody seriously asked which of the protagonists was "the best": the answer depended on what you were looking for in a character. With open weights it is the same for me: the answer is not universal, it is situational. And since the series helped me remember who was who on the API, here is the mapping. It is a cheat sheet, not a benchmark; the real map is still the usage you give it.

## The mapping

| Knight | Provider | Trait to remember |
|---|---|---|
| Seiya (Pegasus) | Qwen | Versatile in every size; good at everything, standout at nothing |
| Shiryu (Dragon) | GLM | From text-only to truly multimodal |
| Hyoga (Cygnus) | Kimi | Cold and precise in code; expensive to keep running |
| Ikki (Phoenix) | DeepSeek | Disruptive arrivals: reasoning, cache and prices |
| Shun (Andromeda) | MiniMax | The apparent weak one who sometimes surprises |

## Seiya (Pegasus Knight) → Qwen

Seiya is the protagonist of the series, but he does not stand out at anything in particular: he is in every arc without being the specialist of any. Qwen works in a similar way. The family covers a huge range of sizes, from models meant to run on modest machines to data-centre ones, with general, code and multimodal checkpoints. If your machine cannot handle the big models, the reasonable bet is Qwen3.8-27B: truly small, downloadable, Apache 2.0 licensed and enough to try things. It is rarely the best on any single axis, but it works for almost everything.

A practical warning: with so many versions, the name does not always tell you what you are running. The published open weights are those of **Qwen3.8-Flash-Next** (125B total parameters, 6B active), while the production model the API serves as `qwen3.8-flash` ships with a one-million-token context by default and built-in tools. Alias and weights are not exactly the same model: read the real ID before trusting the name.

## Shiryu (Dragon Knight) → GLM

In the Sanctuary arc, Shiryu blinds himself so he cannot look at the Medusa Shield of Algol of Perseus, which petrified whoever looked at it. His sight comes back when he awakens the seventh sense in that same arc, he loses it again against Krishna of Chrysaor in the Poseidon arc, and he does not get it back for good until the Hades arc, when he awakens the eighth sense. That first scene, the voluntary sacrifice, is the one that marked me and explains why he was my favourite.

GLM went through a similar cycle. The earlier versions of the GLM-5 series were not multimodal, and **GLM-5.3-Flash** (late August 2026) arrived as the first natively multimodal model of the series (text and image), with MIT weights on Hugging Face, 320B total parameters, 18B active and one million tokens of context. Two caveats: it is not a small model, and the big GLM-5.3 (the *flagship*) is not multimodal. The recovery of "sight", in any case, is the trait that sets it apart.

## Hyoga (Cygnus Knight) → Kimi

Hyoga, Camus's disciple, fights with ice: cold and precise. Kimi gives me a similar feeling: reliable at code, but expensive to keep at full throttle. Cursor based its Composer 2 and 2.5 on Kimi K2.5 (in Composer 2.5, 85% of the compute is Cursor's own training). **Kimi K2.7 Code** (June 12, 2026) is much cheaper and I have used it for infrastructure tasks. **Kimi K3** (API on July 16, 2026, weights on July 27, under its own license that is not MIT) is expensive and not an everyday tool: I save it for when I have quota to spare right before my OpenCode Go usage limit resets, and I use it to run code reviews looking for security vulnerabilities.

## Ikki (Phoenix Knight) → DeepSeek

Ikki has a habit of showing up at exactly the right moment and turning the fight around. DeepSeek does the same with its models: it was among the first to put reasoning on the map for open weights, and it has kept moving the board with price and cache changes since. On September 10, 2026 it shipped **V4.1 Flash**, with MIT weights, and the old aliases (`deepseek-v4-flash`) started pointing to the new version. Its cache in off-peak hours (the lower-demand windows, at a reduced rate; at peak it costs double) sits at $0.003 per million tokens, far below the competition. In an agentic loop, where context comes back again and again, that can decide the bill more than the input price.

## Shun (Andromeda Knight) → MiniMax

Shun looks like the weakest because he is a pacifist: he avoids fighting whenever he can and uses his chain more to defend than to attack, but when things get serious, he delivers. In my experience, MiniMax is one of the providers that gets mentioned the least, but it can be very useful in specific contexts; my impression is that above all in refactors with lots of context. The data at least does not contradict it: **MiniMax M3** (June 1, 2026) ships a one-million-token window (with a guaranteed minimum of 512K on the API). It is not a small model: 428B total parameters and about 23B active. It is natively multimodal, uses a sparse attention architecture ([MSA](https://arxiv.org/abs/2606.13392)) and comes with its own license, not MIT. And it is cheap on top of that.

For refactors or log analysis with lots of context, MiniMax is the Andromeda Chain: the one that shows up out of nowhere and holds everything together.

## Closing

This is a cheat sheet, not a benchmark. If you want numbers, the [quick guide](/en/blog/018-open-weight-models-api/) and the post on [cache and runtimes](/en/blog/019-open-models-cache-agents/) are there. All I wanted here was a map that fits in your head.

*Note: Saint Seiya names and characters are the property of Toei Animation and Masami Kurumada. This post is an original analogy; it does not use images or protected material.*