Model guide

Which model should you run?

GIDE is model-agnostic in a literal sense: the bundled runtime loads roughly 115 architectures, and the agent drives all of them through the same plan gate, write gate and audit log. What changes between models is how they talk to tools, and that is what this page is about.

Start here

Pick by the job, not the leaderboard

Four situations cover almost everything people actually do. The harness is identical in each, only the engine changes.

9B is the floor for coding.Anything smaller will load, hold a conversation and answer questions about your code. It will not reliably carry a multi-step editing loop: small models drift out of their own tool format under pressure, and the agent spends its turns recovering instead of working.
on your machine

Everyday agentic coding

Qwen3.6 · Ornith 9B · Mistral 3, at 9B or above

Tier 1 families at or above the size floor: the pipeline speaks their own tool dialect, so the loop runs grammar-off and the model stays in the format it was trained on.

your own key

Hard planning and refactors

Claude · GPT · Gemini, on your key

Any frontier model works. Plan the difficult turn with one, then hand the mechanical loop back to a local model so the long tail of edits costs nothing.

on your machine

Air-gapped or regulated work

Ornith 9B, bundled

Ships inside the app and runs on first launch with no network at all. Every edit still passes the write gate and lands in the local audit log.

on your machine

The best local output you can fit

GLM-5.x · DeepSeek V4 · gpt-oss 20B

If the memory is there, the larger tier 1 families are the strongest local coders. Check the memory matrix before committing to a download.

Local · tier one

Models that speak tools natively

Their GGUF template carries a tool format the pipeline recognises, so the agent runs grammar-off and the model answers in the dialect it was trained on. This is the strongest local path.

Qwen

Qwen2.5 · Qwen3 · Qwen3.5 / 3.6 (+MoE) · Qwen3-Next · QwQ · Qwen-Coder

dialect: hermes

M

Mistral

Mistral Small / Nemo · Mistral 3 (incl. Medium 3.5) · Large 3 · Mixtral · Ministral

dialect: [TOOL_CALLS]

D

DeepSeek

V3 / V3.x · V4 (incl. V4-Pro) · R1 (full MoE) · R1-Distills

dialect: deepseek markers

K

Kimi

Kimi K2 · K2.x · Kimi-Linear

dialect: deepseek markers

gpt-oss

gpt-oss 20B · gpt-oss 120B

dialect: harmony

Seed-OSS

Seed-OSS · Seed-Coder · Functionary

dialect: hermes

H

Hermes fine-tunes

Nous-Hermes · OpenHermes · most modern agent fine-tunes

dialect: hermes

G

GLM & Command-R

GLM-4 / 4.5 / 4.6 / 5.x (+MoE) · Command-R / R7B / Command-A

dialect: native or clean fallback, depending on the quant's template

One caveat worth stating: “native” means the pipeline speaks their format. A 7–8B model can still drift out of its own format under pressure. That is what the fallback chain exists for, and why output quality tracks the model, not the harness.

Local · tier two

Models the grammar drives

No native tool format in the template, so tool calls go through an enforced-JSON grammar instead. They still run the agent. The dialect is simply ours rather than theirs.

Gemma

Gemma 2 / 3 / 3n / 4 (+MoE)

P

Phi

Phi 2 / 3 / 3.5 / 4 (+MoE)

I

Granite

Granite 3.x · 4.0 / 4.1, including the MoE and hybrid variants

A

OLMo

OLMo · OLMo-2 · OLMoE

Nemotron

Nemotron · Nemotron-H (+MoE)

EXAONE

EXAONE 3 / 4 (+MoE)

OpenELM

OpenELM

ERNIE

ERNIE 4.5

T

Hunyuan

Hunyuan

Arctic

Snowflake Arctic

DBRX

DBRX

Grok

Grok-1 · Grok-2

And the rest

SmolLM3 · StableLM · InternLM2 · MiniCPM · Falcon / Falcon-H1 · StarCoder 1 / 2 · CodeShell · Yi · Vicuna · TinyLlama · CodeLlama · ChatGLM · Baichuan · Jais · Orion · XVERSE · PLaMo 2 / 3 · Arcee · Apertus · dots.llm1 · Ling / Bailing MoE 2 · LFM2 / LFM2-MoE (Liquid) · MiniMax-M2 · BitNet

State-space & hybrid

Mamba · Mamba-2 · Jamba · RWKV-6 / 7

Diffusion language models

LLaDA · LLaDA-MoE · Dream

These load too. That is what “any architecture” is worth in practice.

Online · your own key

Any frontier model works

Connect a provider with your own key and GIDE never sees a copy of it. The key goes to your macOS Keychain, or a permission-restricted file on Windows and Linux, and prompts go straight to the provider you chose.

Anthropic Claude

Claude Opus 5 · Sonnet 5 · Haiku 4.5

Any Claude your key can reach, by model ID

A full Messages-API adapter with tool_use block streaming. Modern Claudes are heavily tool-trained, which makes this the strongest online tool-calling path.

OpenAI GPT

GPT-5.6 Sol · Terra · Luna

Any model ID your key can reach

A Chat Completions adapter that reassembles streamed tool_calls fragments. Works with any model that does function calling on /chat/completions. Models exposed only on the newer Responses API are out of scope.

Google Gemini

Gemini 3.1 Pro · 3.6 Flash · 3.5 Flash-Lite

Any model ID your key can reach

Rides the same OpenAI-shaped adapter through Google's compatibility endpoint, with handling for Gemini's whole-arguments delta quirk.

…and anything else that speaks OpenAI

Point the adapter at any OpenAI-compatible base URL and it works the same way. That is most of the hosted world, and your own servers too.

OpenRouter
hundreds of hosted models behind one key
Together · Fireworks · Groq
hosted Qwen3.6, Kimi K3, GLM-5.2, DeepSeek V4, MiniMax M3
DeepSeek API
V4 and R1, OpenAI-compatible
Mistral La Plateforme
OpenAI-compatible
Azure OpenAI
and enterprise proxies
vLLM · LM Studio · Ollama
your own server, on your own network

The honest part

What it will not do

A support list is only useful if it also says where it stops.

Embedding and encoder models

BERT, nomic-embed, T5-encoder, gemma-embedding. They are in the architecture enum but they are not chat models, so they are no use to an agent.

The vision and audio halves of multimodal models

Qwen-VL, CogVLM, wavtokenizer and friends: the text side may load, but there is no image or audio input. The multimodal weights are dead weight.

Architectures newer than the bundled runtime

The lists above are bounded by llama.cpp b10068, not by what has been released. A point release that reuses its family's architecture loads as soon as the weights exist (Qwen3.6 rides Qwen3.5's, GLM-5.x rides GLM-4's) while a genuinely new architecture has to land in llama.cpp first and reach us on the next node-llama-cpp bump. Anything newer than that is still reachable through a hosted OpenAI-compatible endpoint on your own key.

And one about hardware

Being in the architecture list is not the same as fitting in your memory. The 70B dense models and the very large mixtures (DeepSeek-V3, Grok, DBRX, Arctic) load in principle and are refused in practice by the hardware-floor check, which is the correct outcome. The memory and context matrix has the numbers per machine.

Bring the model. We'll bring the harness.

Plan gate, write gate, memory, tools and the audit log behave identically whichever of these you choose.