Best Models for your Hardware --- 8GB VRAM --- 1. Spark-X2.5-4B:...

Best Models for your Hardware

--- 8GB VRAM ---

1. Spark-X2.5-4B: huggingface.co/XHToken/Spark-…

Based on the benchmarks this model is approaching 9B models in performance, they've focused on coding & agentic performance and the model has a 1M token context window

2. Bonsai-27B: huggingface.co/prism-ml/Terna…

This is the closest you'll get to a frontier model on 8GB, it takes Qwen3.6-27B and quantises the weights to Q2, you get 262K token context window + vision

--- 16GB VRAM ---

1. Gemma-4-12B: huggingface.co/google/gemma-4…
If you're looking for something with vision capabilities this is the top model. It's got so much world knowledge, is great at finding things in images/videos and can maintain coherence over long contexts.

Not the best agent, but seriously holds it's weight.

--- 24GB VRAM ---

TOP PICK

1. Qwen3.8-27B: huggingface.co/Qwen/Qwen3.8-2… I especially recommend Mia Lab's exl3-3.5bpw variant.

I've used this model for coding, game dev, media editing, vllm/sglang config management, fleet management. I'm absolutely shocked at how good it is.

It reasons less, and has MTP so you can run it on decent speeds.

Expect Sonnet/Luna performance

--- 32GB - 64GB ---

TOP PICK

1. Nex-n2.5-Mini: huggingface.co/nex-agi/Nex-N2…

Nex post train strong models like Qwen3.6-35B on agentic/coding/media type tasks, I've tried their models before, they cut down on reasoning lengths and squeeze out a few more IQ points out of the models.

This model is especially amazing for Mac/slower gpus, since it's an MoE and only a small part of the model activates at each token generated.

Expect Sonnet/Luna performance

--- 96GB - 128GB VRAM ---

This is where the frontier really begins.

1. Qwen3.8-Flash-Next: huggingface.co/Qwen/Qwen3.8-F…

Scores 58.7% on deepswe, supports vision, extremely fast, great world knowledge, this can do real work, you won't notice a difference going from Terra/Opus-4.6 to Qwen.

There's 55B params which can be offloaded to cheaper memory or nvme with minimal slowdown, so you'll need:

- amd-strix-halo
- DGX-Spark
- 4x 3090s
- 1x RTX Pro 6000
- Mac m5 ultra/max

----

2. GLM-5.3-Flash-EXL3-2BPW: huggingface.co/0xSero/GLM-5.3…

I designed this to run on 1x DGX Spark or any 128GB system really. It's really good at reverse engineering, and should maintain about 86% the BF16 intelligence.

Still experimental, so please report any possible issues.

--- 196GB - 256gb ---

FRONTIER

1. GLM-5.3-Flash-EXL3-4BPW: huggingface.co/0xSero/GLM-5.3…

91.7% top token agreement, 0.065 kld this is hands down the best model right now for the size range. It'll require some tinkering to get it working right, but IMO this is Opus/Sol level, it's taken everything I've thrown at it, never hit a limit with it.

----

2. DeepSeek-V4-Flash-Vision: huggingface.co/deepseek-ai/De…

DeepSeek is better at conversations IMO, it's more personable and seems to have more world knowledge, excellent agent, not much to say beyond that.

----

3. GLM-5.3-EXL3-3bpw-REAP: huggingface.co/0xSero/GLM-5.3…

This will be the smartest model to run on 2x DGX Sparks/Mac M5-Ultra 256GB for coding, agentic, and reverse engineering. It'll be quite slow, but over night you can use this to RE/Hack/build complex components.

You'll find significant degradation outside of coding/agentic/english so I don't recommend using it for anything outside of those use-cases

----

4. Nex-N2.5-Pro: huggingface.co/nex-agi/Nex-N2…

Incredible model, especially for media editing. Try it

--- 384GB - 512GB ---

1. GLM-5.3: huggingface.co/zai-org/GLM-5.3 I recommend exl3-3bpw to exl3-4bpw for this model, it's really capable of insane development and agentic tasks and goes head to head with Fable/GPT-5.6-Sol

What am i forgetting?

local.ai for recipes (for 128gb or below)