Model Gallery
Discover and install AI models from our curated collection
Find Your Perfect Model
Filter by Model Type
Browse by Tags
lfm2.5-8b-a1b
Try LFM •
Docs •
LEAP •
Discord
# LFM2.5-8B-A1B
LFM2.5 is a new family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.
- **On-device personal assistant**: Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices.
- **Compressed performance**: Competitive with much larger dense and MoE models on instruction following and agentic tasks.
- **Unmatched throughput**: Fastest in its size class on both CPU and GPU inference, with day-one support for llama.cpp, MLX, vLLM, and SGLang.
Find more information about LFM2.5-8B-A1B in our blog post.
**AA-Omniscience Index (higher is better) rewards correct answers and penalizes hallucinations. Scores range from -100 to 100. See more results on Artificial Analysis.*
## 🗒️ Model Details
LFM2.5-8B-A1B is a general-purpose text-only model with the following features:
...
Repository: localaiLicense: other
allenai_olmo-3.1-32b-think
The **Olmo-3.1-32B-Think** model is a large language model (LLM) optimized for efficient inference using quantized versions. It is a quantized version of the original **allenai/Olmo-3.1-32B-Think** model, developed by **bartowski** using the **imatrix** quantization method.
### Key Features:
- **Base Model**: `allenai/Olmo-3.1-32B-Think` (unquantized version).
- **Quantized Versions**: Available in multiple formats (e.g., `Q6_K_L`, `Q4_1`, `bf16`) with varying precision (e.g., Q8_0, Q6_K_L, Q5_K_M). These are derived from the original model using the **imatrix calibration dataset**.
- **Performance**: Optimized for low-memory usage and efficient inference on GPUs/CPUs. Recommended quantization types include `Q6_K_L` (near-perfect quality) or `Q4_K_M` (default, balanced performance).
- **Downloads**: Available via Hugging Face CLI. Split into multiple files if needed for large models.
- **License**: Apache-2.0.
### Recommended Quantization:
- Use `Q6_K_L` for highest quality (near-perfect performance).
- Use `Q4_K_M` for balanced performance and size.
- Avoid lower-quality options (e.g., `Q3_K_S`) unless specific hardware constraints apply.
This model is ideal for deploying on GPUs/CPUs with limited memory, leveraging efficient quantization for practical use cases.
Repository: localaiLicense: apache-2.0
lfm2-1.2b
LFM2-1.2B is a hybrid liquid model designed for edge AI and on-device deployment, offering fast inference and multilingual support across 8 languages. It's optimized for agentic tasks, data extraction, and multi-turn conversations with efficient CPU/GPU/NPU compatibility.
Repository: localaiLicense: lfm1.0
insightface-buffalo-s
Small insightface pack (SCRFD-500MF detector + MBF 512-d embedder +
genderage, ~159MB). Good fit for mid-range CPU deployments.
NON-COMMERCIAL RESEARCH USE ONLY.
Repository: localaiLicense: insightface-non-commercial
insightface-opencv-int8
Int8-quantized OpenCV Zoo face pair (YuNet int8 + SFace int8, ~12MB).
Roughly 3x smaller and noticeably faster on CPU than the fp32 variant
at comparable accuracy for face tasks. APACHE 2.0 — commercial-safe.
Weights are downloaded on install via LocalAI's gallery mechanism.
Repository: localaiLicense: apache-2.0
wespeaker-resnet34
Speaker recognition with WeSpeaker's ResNet34 trained on VoxCeleb,
exported to ONNX. 256-d embeddings, CPU-friendly — avoids the
PyTorch runtime entirely (onnxruntime only). APACHE 2.0.
Pair with the `speaker-recognition` backend's OnnxDirectEngine.
Use when ECAPA-TDNN's torch dependency is undesirable (small
images, edge deployments).
Repository: localaiLicense: cc-by-4.0
rfdetr-cpp-nano
RF-DETR Nano object detection model, served via the native rfdetr.cpp backend (ggml + purego, no Python).
Q8_0 quantization is the recommended default for CPU: same accuracy as F16/F32, ~20MB on disk, fastest CPU latency.
Pure C++/ggml runtime; no Python dependencies. Drop-in for the /v1/detection endpoint.
Repository: localaiLicense: apache-2.0
locate-anything-3b
NVIDIA LocateAnything-3B open-vocabulary object detection (visual grounding), served via the native
locate-anything.cpp backend (C++/ggml + purego, no Python). Describe what to find in a text prompt and
get labeled boxes back; separate multiple categories with . Q8_0 is the recommended default:
box-identical to F16/F32, ~6.3GB, fastest CPU latency. Drop-in for the /v1/detection endpoint (pass the
prompt).
Repository: localaiLicense: other
depth-anything-3-base
Depth Anything 3 (base) monocular metric depth + camera pose, served via the
native depth-anything.cpp backend (C++/ggml + purego, no Python at inference).
Given an image it returns a dense depth map plus the recovered camera
extrinsics (3x4) and intrinsics (3x3). Use GenerateImage (src -> normalized
depth PNG at dst) or Predict (JSON depth stats + pose). q4_k is the
recommended CPU default.
Repository: localaiLicense: apache-2.0
depth-anything-3-small
Depth Anything 3 (small / vits), f32 — the smallest backbone (~131 MB) for
fast CPU depth + camera pose. Same output as base at lower latency.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-base
RF-DETR Base object detection model, served via the native rfdetr.cpp backend.
F16 quantization is recommended on CPU: identical accuracy to F32, half the size, fastest.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-small
RF-DETR Small object detection model (DINOv2-small backbone, 512px input, 3 decoder layers), served
via the native rfdetr.cpp backend (ggml + purego, no Python). A step up from Nano in accuracy while
staying lightweight on CPU. F16 quantization is the recommended default: identical accuracy to F32
at roughly half the size. Drop-in for the /v1/detection endpoint.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-medium
RF-DETR Medium object detection model (DINOv2-small backbone, 576px input, 4 decoder layers), served
via the native rfdetr.cpp backend. Balanced detection quality vs. CPU latency — recommended when
Base is not accurate enough but Large is too slow. F16 quantization is the recommended default:
identical accuracy to F32, half the size. Drop-in for the /v1/detection endpoint.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-large
RF-DETR Large object detection model (DINOv2-small backbone, 704px input, 4 decoder layers), served
via the native rfdetr.cpp backend. Highest-accuracy detection variant — best for offline workflows
and high-resolution inputs where CPU latency is secondary to recall. F16 quantization is the
recommended default: identical accuracy to F32, half the size. Drop-in for the /v1/detection endpoint.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-seg-nano
RF-DETR Seg-Nano instance segmentation model (DINOv2-small backbone, 312px input, 4 decoder layers,
100 queries), served via the native rfdetr.cpp backend. Smallest segmentation variant — fastest CPU
latency, ideal for edge deployment. Returns both bounding boxes and per-instance masks via the
/v1/detection endpoint. F16 quantization is the recommended default: identical accuracy to F32,
half the size.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-seg-small
RF-DETR Seg-Small instance segmentation model (DINOv2-small backbone, 384px input, 4 decoder layers,
100 queries), served via the native rfdetr.cpp backend. Step up from Seg-Nano in mask quality while
staying CPU-friendly. Returns both bounding boxes and per-instance masks via the /v1/detection
endpoint. F16 quantization is the recommended default: identical accuracy to F32, half the size.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-seg-medium
RF-DETR Seg-Medium instance segmentation model (DINOv2-small backbone, 432px input, 5 decoder layers,
200 queries), served via the native rfdetr.cpp backend. Balanced segmentation quality vs. CPU latency
— recommended for everyday segmentation workloads. Returns both bounding boxes and per-instance masks
via the /v1/detection endpoint. F16 quantization is the recommended default.
Repository: localaiLicense: apache-2.0
rfdetr-cpp-seg-2xlarge
RF-DETR Seg-2XLarge instance segmentation model (DINOv2-small backbone, 768px input, 6 decoder layers,
300 queries), served via the native rfdetr.cpp backend. Highest-accuracy segmentation variant — best
for offline workflows and high-resolution inputs where CPU latency is secondary to mask quality.
Returns both bounding boxes and per-instance masks via the /v1/detection endpoint. F16 quantization
is the recommended default: identical accuracy to F32, half the size.
Repository: localaiLicense: apache-2.0
qwen3-30b-a1.5b-high-speed
This repo contains the full precision source code, in "safe tensors" format to generate GGUFs, GPTQ, EXL2, AWQ, HQQ and other formats. The source code can also be used directly.
This is a simple "finetune" of the Qwen's "Qwen 30B-A3B" (MOE) model, setting the experts in use from 8 to 4 (out of 128 experts).
This method close to doubles the speed of the model and uses 1.5B (of 30B) parameters instead of 3B (of 30B) parameters. Depending on the application you may want to use the regular model ("30B-A3B"), and use this model for simpler use case(s) although I did not notice any loss of function during routine (but not extensive) testing.
Example generation (Q4KS, CPU) at the bottom of this page using 4 experts / this model.
More complex use cases may benefit from using the normal version.
For reference:
Cpu only operation Q4KS (windows 11) jumps from 12 t/s to 23 t/s.
GPU performance IQ3S jumps from 75 t/s to over 125 t/s. (low to mid level card)
Context size: 32K + 8K for output (40k total)
Repository: localaiLicense: apache-2.0
qwen3-55b-a3b-total-recall-deep-40x
WARNING: MADNESS - UN HINGED and... NSFW. Vivid prose. INTENSE. Visceral Details. Violence. HORROR. GORE. Swearing. UNCENSORED... humor, romance, fun.
Qwen3-55B-A3B-TOTAL-RECALL-Deep-40X-GGUF
A highly experimental model ("tamer" versions below) based on Qwen3-30B-A3B (MOE, 128 experts, 8 activated), with Brainstorm 40X (by DavidAU - details at bottom of this page).
These modifications blow the model (V1) out to 87 layers, 1046 tensors and 55B parameters.
Note that some versions are smaller than this, with fewer layers/tensors and smaller parameter counts.
The adapter extensively alters performance, reasoning and output generation.
Exceptional changes in creative, prose and general performance.
Regens of the same prompt - even with the same settings - will be very different.
THREE example generations below - creative (generated with Q3_K_M, V1 model).
ONE example generation (#4) - non creative (generated with Q3_K_M, V1 model).
You can run this model on CPU and/or GPU due to unique model construction, size of experts and total activated experts at 3B parameters (8 experts), which translates into roughly almost 6B parameters in this version.
Two quants uploaded for testing: Q3_K_M, Q4_K_M
V3, V4 and V5 are also available in these two quants.
V2 and V6 in Q3_k_m only; as are: V 1.3, 1.4, 1.5, 1.7 and V7 (newest)
NOTE: V2 and up are from source model 2, V1 and 1.3,1.4,1.5,1.7 are from source model 1.
Repository: localaiLicense: apache-2.0
qwen3-22b-a3b-the-harley-quinn
WARNING: MADNESS - UN HINGED and... NSFW. Vivid prose. INTENSE. Visceral Details. Violence. HORROR. GORE. Swearing. UNCENSORED... humor, romance, fun.
Qwen3-22B-A3B-The-Harley-Quinn
This repo contains the full precision source code, in "safe tensors" format to generate GGUFs, GPTQ, EXL2, AWQ, HQQ and other formats. The source code can also be used directly.
ABOUT:
A stranger, yet radically different version of Kalmaze's "Qwen/Qwen3-16B-A3B" with the experts pruned to 64 (from 128, the Qwen 3 30B-A3B version) and then I added 19 layers expanding (Brainstorm 20x by DavidAU info at bottom of this page) the model to 22B total parameters.
The goal: slightly alter the model, to address some odd creative thinking and output choices.
Then... Harley Quinn showed up, and then it was a party!
A wild, out of control (sometimes) but never boring party.
Please note that the modifications affect the entire model operation; roughly I adjusted the model to think a little "deeper" and "ponder" a bit - but this is a very rough description.
That being said, reasoning and output generation will be altered regardless of your use case(s).
These modifications pushes Qwen's model to the absolute limit for creative use cases.
Detail, vividiness, and creativity all get a boost.
Prose (all) will also be very different from "default" Qwen3.
Likewise, regen(s) of the same prompt - even at the same settings - will create very different version(s) too.
The Brainstrom 20x has also lightly de-censored the model under some conditions.
However, this model can be prone to bouts of madness.
It will not always behave, and it will sometimes go -wildly- off script.
See 4 examples below.
Model retains full reasoning, and output generation of a Qwen3 MOE ; but has not been tested for "non-creative" use cases.
Model is set with Qwen's default config:
40 k context
8 of 64 experts activated.
Chatml OR Jinja Template (embedded)
Four example generations below.
IMPORTANT:
See usage guide / repo below to get the most out of this model, as settings are very specific.
If not set correctly, this model will not work the way it should.
Critical settings:
Chatml or Jinja Template (embedded, but updated version at repo below)
Rep pen of 1.01 or 1.02 ; higher (1.04, 1.05) will result in "Harley Mode".
Temp range of .6 to 1.2. ; higher you may need to prompt the model to "output" after thinking.
Experts set at 8-10 ; higher will result in "odder" output BUT it might be better.
That being said, "Harley Quinn" may make her presence known at any moment.
USAGE GUIDE:
Please refer to this model card for
Specific usage, suggested settings, changing ACTIVE EXPERTS, templates, settings and the like:
How to maximize this model in "uncensored" form, with specific notes on "abliterated" models.
Rep pen / temp settings specific to getting the model to perform strongly.
https://huggingface.co/DavidAU/Qwen3-18B-A3B-Stranger-Thoughts-Abliterated-Uncensored-GGUF
GGUF / QUANTS / SPECIAL SHOUTOUT:
Special thanks to team Mradermacher for making the quants!
https://huggingface.co/mradermacher/Qwen3-22B-A3B-The-Harley-Quinn-GGUF
KNOWN ISSUES:
Model may "mis-capitalize" word(s) - lowercase, where uppercase should be - from time to time.
Model may add extra space from time to time before a word.
Incorrect template and/or settings will result in a drop in performance / poor performance.
Can rant at the end / repeat. Most of the time it will stop on its own.
Looking for the Abliterated / Uncensored version?
https://huggingface.co/DavidAU/Qwen3-23B-A3B-The-Harley-Quinn-PUDDIN-Abliterated-Uncensored
In some cases this "abliterated/uncensored" version may work better than this version.
EXAMPLES
Standard system prompt, rep pen 1.01-1.02, topk 100, topp .95, minp .05, rep pen range 64.
Tested in LMStudio, quant Q4KS, GPU (CPU output will differ slightly).
As this is the mid range quant, expected better results from higher quants and/or with more experts activated to be better.
NOTE: Some formatting lost on copy/paste.
WARNING: NSFW. Vivid prose. INTENSE. Visceral Details. Violence. HORROR. GORE. Swearing. UNCENSORED... humor, romance, fun.
Repository: localaiLicense: apache-2.0
Page
1
of many