You could always run a DeepSeek distill on your own machine. That's old news, and the small versions were never the interesting ones. What's new is that the full DeepSeek-V4-Flash, 284 billion parameters, now decodes at 22 to 25 tokens per second on a desktop with a single RTX card and plenty of system RAM, the kind of machine that sits in gaming rooms and offices alike. The tool doing it is called FreeToken, released two weeks ago by a team from Berkeley and UT Austin, free and open source, and any RTX 30, 40 or 50 series card is supported, a 3090 from 2020 included. Push it further and it runs GLM-5.2, a 753 billion parameter model, on a single workstation GPU, and a 35B mixture-of-experts at usable speed on an 8GB laptop card.
If those numbers hold up (they're the authors' own, more on that below), the VRAM wall between frontier-scale models and hardware you can actually buy just moved, and running the new generation of big-but-sparse open models at home stops being a stunt.
Since then my feed has been full of "FreeToken vs Ollama" posts, and most of them miss the point: these tools don't compete. Ollama serves models, LM Studio is a GUI, Unsloth Studio trains models, and FreeToken is a new engine for one specific, difficult job.
I care because of Kimi K3. A month ago I downloaded its weights purely to put them on a shelf: the first open-weight model to crack the frontier top 3, and running it at home meant a 64-GPU supernode and a word with the power company. That's where the big open models have been heading all year, free to download and impossible to run. FreeToken is the first tool that pushes in the other direction, so let's sort out what it actually is, and what it isn't.
Four tools, four jobs
Ollama is the developer default: one command and a model is running behind an OpenAI-compatible API on your machine. MIT-licensed, 88 million in venture money, a claimed 8.9 million monthly developers served by 14 employees, and now a paid cloud tier for the models your GPU can't hold. It runs models. It doesn't train them, and it isn't even the fastest way to run them.
LM Studio is the desktop app: browse Hugging Face, download, chat, serve. Closed source, but free for commercial use since mid-2025. On Apple Silicon its MLX engine is genuinely quick: an independent benchmark this spring measured 102 tokens per second where Ollama got 70 on the same M4 Pro. It's the one I'd install for a non-developer.
Unsloth Studio is the odd one out, and to me the most interesting job of the four: fine-tuning without code. Pick a base model, drop in a dataset, watch the loss curve, export a GGUF. In beta since March, from the two-brother YC startup whose library gets 8 million downloads a month on Hugging Face. Training needs an NVIDIA card; a 4090 is a comfortable floor for an 8B model.
FreeToken is a research engine for one problem: mixture-of-experts models far bigger than your VRAM. NVIDIA-only, Linux and Windows, Apache 2.0, about two weeks old. The author list is what made me sit up: Ion Stoica and Matei Zaharia, the people behind Spark, Ray and vLLM.
Comparing FreeToken to Ollama is comparing a prototype race engine to a delivery van. Both contain an engine. That's about where the overlap ends.
Side by side
| Ollama | LM Studio | Unsloth Studio | FreeToken | |
|---|---|---|---|---|
| The job | serve models locally | GUI + local server | no-code fine-tuning | run huge MoE models |
| Since | 2023 | 2023 | March 2026 | August 2026 |
| License | MIT | proprietary, free for work | Apache 2.0 core, AGPL UI | Apache 2.0 |
| Platforms | Win/Mac/Linux | Win/Mac/Linux | Win/Linux, partial Mac | Win/Linux, NVIDIA only |
| Fine-tunes? | no | no | yes | no |
| Weak spot | slower than raw llama.cpp | closed source | beta, no Mac training | young, unverified numbers |
What FreeToken does differently
Mixture-of-experts models only activate a few experts per token, so the compute is light but every expert still has to be reachable in memory. FreeToken profiles your machine's PCIe and CPU bandwidth on first run, then splits the work between GPU, CPU and system RAM in the ratio your specific box can sustain. No quality tricks either: expert outputs stay bit-exact.
Against the tools everyone uses today, the claimed gains are real: GLM-5.2, the 753B one, decodes at about 15 tokens per second on a single 96GB GPU where llama.cpp manages 7, and Qwen3.6-35B hits 39.3 on an 8GB laptop GPU, about 92% of what a desktop 4090 gets.
The number I find most interesting is none of those. It's tail latency: their worst single turn stayed under 44 seconds, against roughly four minutes for llama.cpp and three for Ollama. If you point a coding agent at a local model, that decides everything, because agents kill turns that hang. FreeToken even ships a command that wires it straight into Claude Code and its cousins.
Which one is for you
If you write code: Ollama stays the default backend, mostly because everything integrates with it. If you want the last 10 to 25% of speed, go one level down to llama.cpp itself.
If you just want to chat with a model, or you're on a Mac: LM Studio. Free for work now, fast on Apple Silicon, and FreeToken doesn't run on Macs at all.
If you want a model that knows your data: Unsloth Studio is the only one of the four that does this. Fine-tuning used to be a week of Python; it's now an evening of clicking. For legal or medical data that can't go near an API, that's a real unlock.
If you own a big NVIDIA card and run a coding agent all day: FreeToken is the one to watch. It's the only tool here built to keep a frontier-scale model running on your own desk all day.
The honest caveats
Every FreeToken number above was measured by its own authors, and the project is two weeks old, so nobody independent has confirmed them yet. One headline claim already doesn't survive a close read: the paper's "2.4x faster" compares an end-to-end figure against pure decode numbers, and like-for-like it's closer to 1.3x. On small dense models llama.cpp is already at par. And there's no Mac support, no multi-GPU, no GGUF, and a core team of about three people.
The others have their own asterisks. Unsloth Studio is beta, and its UI is AGPL, which matters if you want to embed it in a commercial product. Ollama forked its own engine away from llama.cpp and reintroduced bugs upstream had already fixed, after spending a year not crediting llama.cpp at all. LM Studio is closed source, which either bothers you or doesn't.
Where I land
My hot take, held loosely: FreeToken the product probably doesn't win, FreeToken the ideas do. Open source has a habit of folding clever tricks back into the tools everyone already uses, and I'd expect llama.cpp to grow bandwidth-aware expert handling within a year. That's fine. The ideas being free is the point, and the name says as much: the whole pitch is tokens you don't pay for, on hardware you already own.
In July I wrote that the open weights had caught up to the paid flagships, and that the only thing keeping them out of our living rooms was the hardware bill. That bill didn't vanish this month, but it got smaller in a way I didn't expect this soon. The K3 copy on my shelf might boot yet. Future me and the power company are both a little more optimistic this week.
Drawn from the FreeToken paper (arXiv 2608.16157) and GitHub repo, Ollama's Series B announcement and TechCrunch coverage, the LM Studio changelog, Unsloth's documentation, an asiai.dev Apple Silicon benchmark, and a skeptical Cloud Codes analysis of the paper's framing. A snapshot as of August 25, 2026. Every FreeToken benchmark is author-measured and unverified at publication, and star counts, prices and version numbers age fast, so check the repos before you quote me.