A model is not an app.

Hermes 4, Qwen and Llama are weights: files of numbers. They have no window, no memory, no access to your files and no way to run anything. The software around them is a separate thing, and you can get that software free from several places. This page explains the difference rather than selling you one.

Verified on 31 July 2026

The short answer

If you want an open-weight model working on first launch, with no terminal, Kynd Agent does the picking, quantising, serving and tool-wiring for you. It sizes a model to your RAM, benchmarks your Mac, and runs it through MLX. That convenience is the whole of what the $79 buys. It is not better than the weights; it is a different layer.

If you are happy choosing your own weights and running your own server, you do not need to pay anyone. Ollama is free and MIT-licensed. LM Studio is free, supports MLX, and its own announcement says it is free to use at work as well as at home. Jan is Apache 2.0 and free with no paid tier at all. And the weights themselves are free forever.

The category

Weights are not software.

What you download from Hugging Face is a model's parameters, a config and a tokenizer. Nous publishes Hermes in BF16 safetensors and in GGUF. That is the entire delivery.

A model on its own does one thing: given some text, it predicts more text. Everything else you associate with an AI assistant is the application wrapped around it.

No interface. There is no window to open. Something else has to draw one.
No memory. A model does not remember yesterday, or the start of a long conversation, unless an app stores that and feeds it back in.
No file access. Weights cannot open, read or write anything on your disk.
No agent loop. Nothing plans, retries, or checks its own work. That loop is code somebody wrote.
Tool calls are only a suggestion. See below — this is the part people most often misread.

When a model "uses a tool", it does not use anything. It emits text in an agreed shape and hopes something is listening. Hermes wraps its tool calls in <tool_call> tags; a surrounding program has to notice that text, parse it, run the thing it describes, and hand the result back. If nothing is listening, the model has written a note nobody reads.

That is the whole model-versus-app distinction, and it is why "which is better, Hermes 4 or an app?" is not a question with an answer. One is an engine. The other is a car.

A name that trips people up

Two different things are called Hermes.

Nous Research publishes both, and they are not bundled. Hermes Agent is an app — free, MIT-licensed, cross-platform, and it genuinely edits files and runs commands. Hermes 4 and Hermes 4.3 are models — weights on Hugging Face.

Hermes Agent ships no model and no inference engine of its own. After you install it you still have to tell it which provider to use, and its own FAQ names OpenRouter and the Nous Portal as recommended providers. So if you have been comparing "Hermes" against something, work out which of the two you mean first. We wrote about the app separately.

The actual work

What running one takes.

Six steps, none of them mysterious, all of them yours. The specifics here are from the Hermes model cards, checked on 31 July 2026.

Pick a size and a quantisation. The current flagship, Hermes 4.3 36B, needs roughly 20 GB or more for the weights alone at 4-bit, before any KV cache. On a 16 GB Mac, realistically only the 14B fits.
Download multi-gigabyte files. BF16 safetensors or GGUF, depending on which runtime you have settled on.
Install a runtime. The official model-card guidance covers Transformers, vLLM and SGLang. There is no macOS section and no MLX section on any official Hermes model card, and Nous publishes no official MLX build — the MLX quants you will find are third-party community uploads, and nobody official has verified their quality.
Wire up tool-call parsing. The card tells you to set the tool parser to hermes in vLLM or qwen25 in SGLang. Neither is what a Mac user is typically running, so on llama.cpp or Ollama you handle the parsing yourself.
Turn on reasoning mode. It needs a thinking=True template flag, or a specific long system prompt. The model emits its reasoning in <think> tags.
Check the licence. Verified on Hugging Face, 31 July 2026: Hermes 4.3 36B and Hermes 4 14B are apache-2.0, while Hermes 4 70B and 405B carry Meta's llama3 licence. If you plan to redistribute or commercialise, that difference matters.

Qwen and Llama turn up in this story too: Hermes 4 14B is built on Qwen3-14B, and the 70B and 405B are built on Meta's Llama 3.1, which is where the llama3 licence comes from. We have only verified the Hermes cards, so if you are running a different family, read its own card rather than trusting a summary of ours.

One thing we will not give you: speed figures. No one official has published Apple Silicon performance numbers for these models. Anyone quoting you tokens per second on a Mac, ourselves included, is quoting something they measured on their own machine.

None of the six steps is hard for someone who enjoys this sort of thing. It is a couple of evenings for someone who does not.

Side by side

Three layers, not three rivals.

	Kynd Agent	Open weights on their own	A runtime or model app
What it is	An app with a model inside it	Parameters, a config and a tokenizer	Software that loads weights and serves them
What you download	One 919 MB app, which then fetches a model sized to your RAM	Multi-gigabyte weight files — BF16 safetensors or GGUF for Hermes	An installer, then weights from an in-app catalogue
Interface	Chat, Deep Research and a builder	None	Ollama and LM Studio both ship a desktop chat app; LM Studio also runs a local server
Edits files, runs commands	Yes	No — it can only emit tool-call JSON for something else to run	Ollama's desktop app: not documented. LM Studio's separate Bionic app: file, search, Git and shell tools
Tool-call parsing	Handled inside the app	Yours to wire up; documented parsers exist for vLLM and SGLang only	Depends on the runtime
Apple Silicon	MLX	No official MLX build from Nous; community quants only	LM Studio supports MLX; Ollama's MLX support is in preview
Who picks the model	The app benchmarks your Mac and picks	You do	You do, from a catalogue
Price	$79 founding price (then $149), 30-day trial, free tier forever	Free. apache-2.0 on Hermes 4.3 36B and 4 14B; llama3 on 70B and 405B	Ollama free and MIT. LM Studio free, including at work. Jan free and Apache 2.0

Each one, fairly

The pieces, described plainly.

Hermes 4 and Hermes 4.3

Free weights · Hugging Face · licence varies by size

An open-weight family from Nous Research. Genuinely open: the two newest sizes are apache-2.0, so you can inspect them, fine-tune them and redistribute them in ways no closed model allows. They are deliberately low-refusal and steerable, which is a real reason to want these particular weights if over-refusal is your complaint about mainstream assistants. They do hybrid reasoning with explicit <think> tags, and function calling is trained in rather than bolted on. Free to download and free to run, forever.

They are not a product you compare to an app. There is no interface, no memory, no file library, no voice, no research agent, no coder — all of that would have to be built around them. There is no official Mac or Apple Silicon support of any kind, and no vision variant in the family, so it is text only. The 4.3 36B dates from November 2025.

Ollama

Free, MIT · macOS 14+, Windows, Linux · local, with optional cloud

The de-facto open-source runtime for open models, now with an official desktop chat app. Install is a DMG drag-and-drop; a terminal one-liner is offered as an alternative, not a requirement. Models download from inside the app, from a very large curated library. It is MLX-powered on Apple Silicon, in preview. Running models on your own hardware is, in their words, always unlimited. Verified on 31 July 2026: 177,409 stars on GitHub, pushed the same day — by far the largest ecosystem here, and the backend dozens of other tools plug into.

The desktop app is chat. It reasons over files you drag in, but no file editing, command execution or planning agent is documented, and there is no documented persistent memory feature. Its web search is a hosted API needing a free account and a key, not something local. Agentic behaviour means pointing a separate third-party tool at it. Many people will run Ollama alongside something else rather than instead of it.

LM Studio, and Bionic

Free · macOS 14+ Apple Silicon, Windows, Linux · local, with optional cloud

A desktop app for running open models locally, with a large in-app catalogue and no terminal needed. It supports MLX natively, so it concedes nothing to Kynd on Apple Silicon. Their own announcement of 8 July 2025 reads: "Starting today, LM Studio is free to use both at home and at work." There is an OpenAI-compatible REST server, MIT-licensed Python and TypeScript SDKs, an MIT-licensed lms CLI, and a headless daemon. Bionic, a separate app launched on 16 July 2026, adds Code Projects "for working in a local codebase with file, search, Git, and shell tools".

The app itself has no published open-source licence — the SDKs and CLI are MIT, the app is not, so it should not be called open source. Bionic is new, so its maturity and roadmap are unproven, and Bionic Pass pricing has not been announced. Voice appears to be input only: local transcription, with no voice output or cloning documented. No persistent cross-session memory, project templates, live preview or visual page editor are documented.

Jan

Free, Apache 2.0 · macOS 13.6+, Windows, Linux · local, with optional cloud

A fully open-source desktop app positioned as an open-source replacement for cloud assistants. The licence is Apache 2.0 — read the LICENSE file in the janhq/jan repository rather than GitHub's API field, which reports NOASSERTION for a benign reason. Free with no paid tier at all. Its stated RAM guidance starts at 8 GB for smaller models, below Kynd's floor. It ships native web search and fetch, exposes a local OpenAI-compatible server on port 1337, and has broad MCP support with per-tool permissions. Verified 31 July 2026: 43,784 stars, pushed the same day.

It is not an agentic coder. It serves models to agents rather than acting on your files itself, and getting agentic behaviour means assembling several pieces yourself. Its MCP tools need Node and/or Python already installed. Memory is listed as "Coming Soon" on their own homepage, so it has not shipped.

Kynd Agent

$79 founding price (then $149) · macOS 13+, Apple Silicon, 16 GB · local

One 919 MB download that benchmarks your Mac, picks an open-weight model that fits your RAM, serves it through MLX and wires up the tool-calling, so the model can hold a conversation with memory, search your own files, read aloud in an on-device voice, run multi-step research you watch happen, and build software with templates, a live preview and a visual page editor. No account, no telemetry, works offline. Thirty-day trial, and a free tier that stays free.

It is the app layer, and only that. It is proprietary where Ollama, Jan and the weights are not. It is Apple Silicon only, where every alternative here runs on Windows and Linux too. It cannot reach a frontier cloud model when a task is beyond a 16 GB machine, and it does not let you bring arbitrary weights of your own choosing — the point is that it chose for you.

Be honest

When to use something else.

Skip the app, or pick a different one.

The weights on their own — you want to fine-tune, inspect or redistribute a model, or you specifically want a low-refusal steerable one. No app can give you that, and it costs nothing.
Ollama — you are comfortable in a terminal, or you want to serve a local model to other software over a clean API. Free, MIT, cross-platform, and the biggest ecosystem on this beat.
LM Studio — you want a free app with a browsable model catalogue, MLX on Apple Silicon, and no purchase to make. Free at work as well as at home, in their own words. If you will not pay for local AI on a Mac, this is the honest answer.
Jan — you want open source you can audit and fork, a local API server, or a machine with 8 GB of RAM that Kynd cannot serve at all.
Zed — you can already read code and want a free, open-source editor whose agent works with llama.cpp, Ollama, LM Studio or any local OpenAI-compatible server, and which auto-discovers models Ollama has already pulled.
Hermes Agent — you want a free, MIT-licensed agent that runs on Windows or Linux as well as a Mac, or headless on a server, and you are happy to configure your own model provider.

Also worth reading

The best local AI apps for MacThe app layer compared: LM Studio, Ollama, Jan, AnythingLLM, Msty and GPT4All. Most of them are free. Hermes Agent, the appThe other thing called Hermes: an agent that asks you for a model, next to one that brings its own. Coding agents that run locallyCursor, Copilot, Cline, Zed and Devin Desktop. Several of them run local models by design.

Try it for thirty days.

No account, no card. Apple Silicon Mac with 16 GB or more.

Download Kynd Agent