Unlocked LM Studio Backends (v1.59.0): AVX1 & More Supported – Testers Wanted

Posted by TheSpicyBoi123@reddit | LocalLLaMA | View on Reddit | 13 comments

Hello everyone!

The latest patched backend versions (1.59.0) are now out, and they bring full support for “unsupported” hardware via a simple patch (see GitHub). Since the last update 3 months ago, these builds have received major refinements in performance, compatibility, and stability via optimized compiler flags and work by llama cpp team.

Here’s the current testing status:

✅ AVX1 CPU builds: working (tested on Ivy Bridge Xeons)
✅ AVX1 Vulkan builds: working (tested on Ivy Bridge Xeons + Tesla K40 GPUs)
❓ AVX1 CUDA builds: untested (no compatible hardware yet)
❓ Non-AVX experimental builds: untested (no compatible hardware yet)

I’m looking for testers to try the newest versions on different hardware, especially non-AVX2 CPUs and newer NVIDIA GPUs, and share performance results. Testers are also wanted for speed comparisons of the new vs old cpu backends.

👉 GitHub link: lmstudio-unlocked-backend

Brief install instructions:
- navigate to backends folder. ex C:\Users\Admin\.lmstudio\extensions\backends
- (recommended for clean install) delete everything except "vendor" folder
- drop contents from compressed backend of your choice

- select it in LM Studio runtimes and enjoy.

Model

Variant

Tokens per second (tok/sec)

Total tokens

Time to first token (s)

Seed36b

1.98

295

4.61

Seed36b

2.39

300

0.93

Seed36b

2.34

295

0.95

Gemma-3 12B projection abliterated

3.07

300

4.13

Gemma-3 12B projection abliterated

4.06

300

0.40

Gemma-3 12B projection abliterated

4.10

300

0.42

[-]

fiery_prometheus@reddit

Tangential question, do you support avx512?

TheSpicyBoi123@reddit (OP)

Hello, and yes! You can build a backend with AVX512 using the generator script. I would be quite curious in your performance on AVX512 using the optimizer levels from none to 03. I can make you a custom one later for you to try if you for some reason have difficulty.

Here are some benchmark results for avx512 with different optimizer flags for MSVC. Also, the backend-manifest.json should have a unique name, otherwise lmstudio doesn't pick up on more backends, and this isn't written in the guide.

I ran benchmarks with both seed36b and gemma-3 12b.

Model	Variant	Tokens per second (tok/sec)	Total tokens	Time to first token (s)
Seed36b	Od	1.98	295	4.61
Seed36b	O2	2.39	300	0.93
Seed36b	Ox	2.34	295	0.95
Gemma-3 12B projection abliterated	Od	3.07	300	4.13
Gemma-3 12B projection abliterated	O2	4.06	300	0.40
Gemma-3 12B projection abliterated	Ox	4.10	300	0.42

I've attached the build script, it builds on windows and makes 3 versions with different optimizer flags.

https://pastebin.com/Hjs3q43Z

I fully agree, the backend manifest needs some cleaner labelling. If you are feeling motivated, I can gladly make you a contributor on the github page too.

Would be neat, always liked the project :-) https://github.com/Nidvogr

Awesome, just added you. I would love your help with cleaning up the build scripts (and have a look at the new ones, I made)

The script is also a lot cleaner then mine, great job lmao