Models
A catalogue of 127 open models you can run on the AETRON network: LLMs, image and video generation, speech and embeddings. Licence, memory requirements and parameters on every entry.
- Qwen2.5-0.5B-Instruct · Alibaba / Qwen
Tiny 0.5B instruction-tuned LLM — runs on edge devices and CPUs.
- Qwen3.5-9B · Alibaba / Qwen
9B dense LLM from the Qwen3.5 generation (02.2026). Strong reasoning, coding and multilingual chat.
- Qwen3.5-9B (FP8-dynamic) · Alibaba / Qwen
Ready-made FP8 build (compressed-tensors): 9B dense LLM from the Qwen3.5 generation (02.2026). Strong reasoning, coding and multilingual chat.
- Qwen3.5-9B (w4a16) · Alibaba / Qwen
Ready-made INT4 W4A16 build (compressed-tensors): 9B dense LLM from the Qwen3.5 generation (02.2026). Strong reasoning, coding and multilingual chat.
- Qwen-Image · Alibaba Qwen
20B MMDiT text-to-image. Best-in-class text rendering, English and Chinese.
- Qwen3.6-35B-A3B · Alibaba / Qwen
MoE 35B with ~3B active: near-27B quality, much faster.
- Qwen3.6-35B-A3B (NVFP4 (compressed-tensors)) · Alibaba / Qwen
Ready-made NVFP4 (compressed-tensors) build - runs on the fleet stack (compressed-tensors). Blackwell window.
- Qwen3.6-35B-A3B (official FP8) · Alibaba / Qwen
Ready-made official FP8 build of Qwen3.6-35B-A3B.
- Qwen3.5-4B · Alibaba / Qwen
Compact 4B for consumer GPUs and edge.
- Qwen3.5-4B (INT4 W4A16) · Alibaba / Qwen
Ready-made INT4 W4A16 build of Qwen3.5-4B.
- Qwen3.5-4B (FP8-dynamic) · Alibaba / Qwen
Ready-made FP8-dynamic build of Qwen3.5-4B.
- Qwen3.5-9B (GGUF Q4_K_M) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q4_K_M) of Qwen3.5-9B - one artifact runs on Metal, CUDA and CPU; the build the GGUF consensus engine was proven on.
- Qwen3.5-9B (GGUF Q4_K_M + vision) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q4_K_M) of Qwen3.5-9B - one artifact runs on Metal, CUDA and CPU; the build the GGUF consensus engine was proven on. Manifest carries the F16 projector, so the entry reads images as well as text.
- Qwen3.5-27B (GGUF Q8_0) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, 8-bit) of Qwen3.5-27B - same model, higher-fidelity quant; a separate entry because a different artifact is a different model_hash.
- Qwen3.8-27B (GGUF UD-Q4_K_M) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_M) of Qwen3.8-27B - one artifact runs on Metal, CUDA and CPU. 262144-token context. The upstream model is multimodal (image-text-to-text); this pin covers the TEXT weights only - the vision projector (mmproj) is a separate artifact and is not part of this manifest.
- Qwen3.8-27B (GGUF UD-Q4_K_M + vision) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_M) of Qwen3.8-27B - one artifact runs on Metal, CUDA and CPU. 262144-token context. The upstream model is multimodal (image-text-to-text); this pin covers the TEXT weights only - the vision projector (mmproj) is a separate artifact and is not part of this manifest. Manifest carries the F16 projector, so the entry reads images as well as text.
- Qwen3-Coder-30B-A3B · Alibaba / Qwen
Coding MoE with 256K context, ~3B active.
- Qwen3-Coder-30B-A3B (official FP8) · Alibaba / Qwen
Ready-made official FP8 build of Qwen3-Coder-30B-A3B.
- Gemma 4 E4B IT · Google
Gemma 4 efficient 4B instruction-tuned.
- Gemma 4 E4B IT (GGUF Q4_0) · Google
llama.cpp GGUF build (Q4_0 QAT) of Gemma 4 E4B IT. One projector file carries BOTH encoders - clip.has_vision_encoder and clip.has_audio_encoder - so the entry reads images and speech on the GGUF consensus engine at 6.1 GB instead of the 18 GB the torch path needs.
- Gemma 4 12B IT · Google
Gemma 4 12B, instruction-tuned: chat with images, 256K context.
- Gemma 4 31B IT · Google
Gemma 4 31B, instruction-tuned: Google's largest dense open model, chat with images.
- Gemma 4 31B (official QAT W4A16) · Google
Ready-made official QAT W4A16 build of Gemma 4 31B.
- Gemma 4 31B (FP8-block) · Google
Ready-made FP8-block build of Gemma 4 31B.
- Ministral 3 8B · Mistral AI
Vision-capable 8B from the edge-focused Ministral 3 line.
- Ministral 3 14B · Mistral AI
Larger Ministral 3 with vision.
- Ministral 3 14B (FP8-dynamic) · Mistral AI
Ready-made FP8-dynamic build of Ministral 3 14B.
- Nemotron-3 Nano 4B · NVIDIA
Hybrid Mamba-2 Nano 4B, long context.
- gpt-oss-20b · OpenAI
OpenAI's 21B MoE in native MXFP4; the miner runs it dequantized to bf16, in 42 GB.
- GLM-4.7-Flash · Z.ai
Consumer GLM 31B.
- GLM-4.7-Flash (FP8-dynamic) · Z.ai
Ready-made FP8-dynamic build of GLM-4.7-Flash.
- DeepSeek-V4-Pro (NVFP4-FP8 (compressed-tensors)) · DeepSeek
Ready-made NVFP4-FP8 (compressed-tensors) build - runs on the fleet stack (compressed-tensors). Blackwell window.
- Qwen3-Coder-Next · Alibaba / Qwen
SOTA open coder, 80B-A3B.
- Qwen3-Coder-Next (official FP8) · Alibaba / Qwen
Ready-made official FP8 build of Qwen3-Coder-Next.
- Nemotron-3 Super 120B-A12B · NVIDIA
Super 120B-A12B: single 80 GB card at 4-bit.
- gpt-oss-120b · OpenAI
OpenAI's 117B MoE in native MXFP4; the miner runs it dequantized to bf16, in 234 GB on several GPUs.
- Mistral Small 4 119B · Mistral AI
Unified instruct+reasoning 119B.
- FLUX.2 [klein] 4B · Black Forest Labs
Distilled consumer FLUX.2, 4B.
- Z-Image Turbo · Tongyi-MAI
6B S3-DiT: best quality-per-GB, ~2s/image on 4090.
- Chroma1-HD · Lodestone Rock
Apache base for unrestricted finetuning.
- Wan 2.2 TI2V-5B · Alibaba / Wan-AI
5B text+image to video. Diffusers mirror: the upstream repo ships its VAE and T5 as pickle, which the loader rejects.
- Wan 2.2 T2V-A14B · Alibaba / Wan-AI
MoE video, 27B total / 14B active. Diffusers mirror; both experts stay resident, so VRAM is for the pair.
- HunyuanVideo 1.5 · Tencent
8.3B video model, 480p t2v. Diffusers mirror: the upstream repo needs Tencent's own hyvideo package and ships no text encoder.
- LTX-2.3 Distilled 1.0 · Lightricks
Audio+video generation, speed champion. Manifest pins the distilled 1.0 checkpoint in the diffusers format (diffusers/LTX-2.3-Distilled-Diffusers).
- Chandra OCR 2 · Datalab
SOTA OCR on Qwen3.5, 40+ languages.
- Qwen3-Embedding-0.6B · Alibaba / Qwen
Text embedding leader line, CPU-capable size.
- Qwen3-Embedding-8B · Alibaba / Qwen
Top MMTEB text embeddings.
- MiniMax H3 (int8, ComfyUI) · MiniMax (ComfyUI weights: Comfy-Org)
MiniMax H3 text-to-video with sound on ComfyUI: Comfy-Org's int8 DiT and text encoder, the 4-step turbo LoRA and both VAEs.
- Qwen3.8-27B · Alibaba / Qwen
Dense 27B Qwen3.8, natively multimodal (text + images).
- Qwen3.8-27B (FP8) · Alibaba / Qwen
Official FP8 build of Qwen3.8-27B.
- Qwen3.8-27B (NVFP4) · Alibaba / Qwen · Unsloth
NVFP4 (Blackwell) build of Qwen3.8-27B by Unsloth, compressed-tensors.
- Nemotron 3.5 Lightning 30B-A3B · NVIDIA
Hybrid Mamba-Transformer MoE, 31.6B total / 3B active.
- Nemotron 3.5 Lightning 30B-A3B (FP8) · NVIDIA / Red Hat AI
FP8 build of Nemotron 3.5 Lightning 30B-A3B by Red Hat AI.
- MiniCPM5-2B · OpenBMB
Compact 2.5B LLM for edge devices and CPUs.
- MiniCPM-V 4.6 · OpenBMB
1.3B vision-language model: images and documents on small GPUs.
- PaddleOCR-VL 1.6 · Baidu / PaddlePaddle
0.96B document-parsing VLM: text, tables, formulas.
- Qwen3.5-0.8B · Alibaba / Qwen
Smallest Qwen3.5: 0.87B, text and images, runs on CPUs.
- DeepSeek-V4-Flash-0731 · DeepSeek
DeepSeek V4 Flash, 31.07.2026 update: 304B MoE (FP8 + FP4 experts).
- DeepSeek-V4-Pro-0813 · DeepSeek
DeepSeek V4 Pro, 13.08.2026 update: 1.65T MoE (FP8 + FP4 experts).
- DeepSeek-V4-Flash-0731 (NVFP4) · DeepSeek / Red Hat AI
NVFP4 build of DeepSeek-V4-Flash-0731 by Red Hat AI, compressed-tensors.
- GLM-5.3 · Z.ai
GLM-5.3: 753B MoE, published in FP8 by Z.AI.
- MiniMax-M3 · MiniMax
MiniMax M3: 427B MoE with vision.
- Qwen3-ASR-1.7B (GGUF Q8_0) · Alibaba / Qwen
Speech recognition, 1.7B, on the llama.cpp engine.
- Qwen-Image-2.1 (GGUF Q8_0, sd.cpp) · Alibaba Qwen (GGUF: leejet; VAE: Comfy-Org)
Text-to-image with Qwen-Image-2.1, 40 steps, on sd.cpp: Q8_0 DiT, Qwen3-VL-8B text encoder, up to 2752 px a side. Research license.
- Qwen-Image-2.1 Edit (GGUF Q8_0, sd.cpp) · Alibaba Qwen (GGUF: leejet; VAE: Comfy-Org)
Image editing with Qwen-Image-2.1, 40 steps, on sd.cpp: the Q8_0 set plus the Qwen3-VL-8B vision projector, up to 10 references. Research license.
- Qwen-Image-2.1 (GGUF Q4_K, sd.cpp) · Alibaba Qwen (GGUF: leejet; VAE: Comfy-Org)
Text-to-image with Qwen-Image-2.1, 40 steps, on sd.cpp: Q4_K DiT for 12-16 GB cards. Research license.
- Qwen-Image-2.1 Turbo (GGUF Q8_0, sd.cpp) · Alibaba Qwen (GGUF: unsloth; VAE: Comfy-Org)
Text-to-image with the official Qwen-Image-2.1-Turbo, 8 steps, on sd.cpp: Q8_0 DiT, Qwen3-VL-8B text encoder. Research license.
- Qwen-Image-2.1 Turbo Edit (GGUF Q8_0, sd.cpp) · Alibaba Qwen (GGUF: unsloth; VAE: Comfy-Org)
Image editing with the official Qwen-Image-2.1-Turbo, 8 steps, on sd.cpp: the Q8_0 set plus the Qwen3-VL-8B vision projector, up to 10 references. Research license.
- Qwen-Image-2.1 Turbo (GGUF Q4_K_M, sd.cpp) · Alibaba Qwen (GGUF: unsloth; VAE: Comfy-Org)
Text-to-image with the official Qwen-Image-2.1-Turbo, 8 steps, on sd.cpp: Q4_K_M DiT for 12-16 GB cards. Research license.
- Qwen3.8-Flash-Next (GGUF UD-Q4_K_XL) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL, Unsloth) of Qwen3.8-Flash-Next, a preview of the Qwen4 architecture: 177B MoE, 512 experts with 10 active, hybrid linear and sparse attention, 262144-token context.
- Qwen3.8-Flash-Next (GGUF UD-Q4_K_XL + vision) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL, Unsloth) of Qwen3.8-Flash-Next, a preview of the Qwen4 architecture: 177B MoE, 512 experts with 10 active, hybrid linear and sparse attention, 262144-token context. The manifest carries the F16 vision projector, so the entry reads images as well as text.
- Qwen3.8-Flash-Next (GGUF Q8_0) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Qwen3.8-Flash-Next, a preview of the Qwen4 architecture: 177B MoE, 512 experts with 10 active, hybrid linear and sparse attention, 262144-token context.
- Qwen3.8-Flash-Next (GGUF Q8_0 + vision) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Qwen3.8-Flash-Next, a preview of the Qwen4 architecture: 177B MoE, 512 experts with 10 active, hybrid linear and sparse attention, 262144-token context. The manifest carries the F16 vision projector, so the entry reads images as well as text.
- Qwen3.8-27B (GGUF Q8_0) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Qwen3.8-27B - one artifact runs on Metal, CUDA and CPU. 262144-token context.
- Qwen3.8-27B (GGUF Q8_0 + vision) · Alibaba / Qwen (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Qwen3.8-27B - one artifact runs on Metal, CUDA and CPU. 262144-token context. The manifest carries the F16 vision projector, so the entry reads images as well as text.
- Nemotron 3.5 Bolt 74B-A9B (GGUF Q8_0) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Nemotron 3.5 Bolt: hybrid Mamba-2 and Transformer MoE, 74B total, 9B active, 256K context.
- Nemotron 3.5 Bolt 74B-A9B (GGUF Q8_0 + vision) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of Nemotron 3.5 Bolt, 74B MoE with 9B active. The BF16 vision projector reads images.
- Nemotron 3.5 Bolt 74B-A9B (GGUF UD-Q4_K_XL) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL) of Nemotron 3.5 Bolt, 74B MoE with 9B active, 256K context.
- Nemotron 3.5 Bolt 74B-A9B (GGUF UD-Q4_K_XL + vision) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL) of Nemotron 3.5 Bolt, 74B MoE with 9B active. The BF16 vision projector reads images.
- Nemotron 3.5 Bolt E4B · NVIDIA
NVIDIA's small model of the 3.5 generation, post-trained from Gemma 4 E4B for tool use and low latency; 8B total, 4B effective.
- Nemotron 3.5 Bolt E4B (NVFP4) · NVIDIA (NVFP4: Red Hat AI)
Ready-made NVFP4 build (compressed-tensors, Red Hat AI) of Nemotron 3.5 Bolt E4B, for Blackwell GPUs.
- Nemotron 3.5 Bolt E4B (GGUF Q4_K_M) · NVIDIA
NVIDIA's own llama.cpp GGUF build (Q4_K_M) of Nemotron 3.5 Bolt E4B: one file for Metal, CUDA, ROCm and CPU.
- Nemotron 3.5 Bolt E4B (GGUF Q8_0) · NVIDIA (GGUF: ggml-org)
llama.cpp GGUF build (Q8_0, ggml-org) of Nemotron 3.5 Bolt E4B: one file for Metal, CUDA, ROCm and CPU.
- Nemotron 3.5 Super VL 120B-A12B (GGUF UD-Q4_K_XL) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL) of Nemotron 3.5 Super VL: Mamba-2 and attention MoE, 120B total, 12B active.
- Nemotron 3.5 Super VL 120B-A12B (GGUF UD-Q4_K_XL + vision) · NVIDIA (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL) of Nemotron 3.5 Super VL, 120B MoE with 12B active. The BF16 projector reads images.
- Nemotron 3.5 Super VL 120B-A12B (GGUF Q8_0 + vision) · NVIDIA (GGUF: ggml-org)
llama.cpp GGUF build (Q8_0, ggml-org) of Nemotron 3.5 Super VL, 120B MoE with 12B active. The BF16 projector reads images.
- MiniCPM-V 4.7 1B (GGUF Q8_0 + vision) · OpenBMB
OpenBMB's own llama.cpp GGUF build (Q8_0) of MiniCPM-V 4.7 1B: a small vision-language model for CPU, Mac and any GPU.
- MiniCPM-V 4.7 35B-A3B (GGUF Q4_K_M + vision) · OpenBMB
OpenBMB's own llama.cpp GGUF build (Q4_K_M) of MiniCPM-V 4.7 35B-A3B: a MoE vision-language model with 3B active, 22 GB of weights.
- MiniCPM-V 4.7 35B-A3B (GGUF Q8_0 + vision) · OpenBMB
OpenBMB's own llama.cpp GGUF build (Q8_0) of MiniCPM-V 4.7 35B-A3B, a MoE vision-language model with 3B active.
- MiniCPM-OCR 1.0 (GGUF Q8_0) · OpenBMB
1.3B document parser (text, tables, formulas, charts) as OpenBMB's own llama.cpp GGUF build: runs on CPU, Mac and any GPU.
- Granite Speech 5.0 2B (GGUF Q8_0) · IBM
IBM's speech recognition and speech translation for English, French, German, Spanish, Portuguese and Japanese, as a llama.cpp GGUF build.
- DeepSeek-V4.1-Flash-1017 (GGUF UD-Q4_K_XL) · DeepSeek (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL, MXFP4 experts) of DeepSeek-V4.1-Flash-1017: 763B MoE.
- DeepSeek-V4.1-Flash-1017 (GGUF UD-Q4_K_XL + vision) · DeepSeek (GGUF: unsloth)
llama.cpp GGUF build (Unsloth Dynamic Q4_K_XL, MXFP4 experts) of DeepSeek-V4.1-Flash-1017, 763B MoE. The BF16 projector reads images.
- Nemotron 3.5 Embed 1B · NVIDIA
Multilingual text embeddings from NVIDIA, 1.6B, mean pooling, L2-normalised vectors.
- Nemotron 3.5 Embed 8B · NVIDIA
Multilingual text embeddings from NVIDIA, 8.4B, mean pooling, L2-normalised vectors.
- GLM-5.3 (NVFP4) · Z.ai (NVFP4: Red Hat AI)
Ready-made NVFP4 build (compressed-tensors, Red Hat AI) of GLM-5.3 for Blackwell GPUs: 465 GB against 756 GB in FP8.
- Qwen3.5-9B (NVFP4) · Alibaba / Qwen (NVFP4: Red Hat AI)
Ready-made NVFP4 build (compressed-tensors, Red Hat AI) of Qwen3.5-9B for Blackwell GPUs.
- Granite Switch 4.1 8B (GGUF Q4_K_M) · IBM
IBM's own llama.cpp GGUF build (Q4_K_M) of Granite Switch 4.1 8B.
- Granite Switch 4.1 8B (GGUF Q8_0) · IBM
IBM's own llama.cpp GGUF build (Q8_0) of Granite Switch 4.1 8B.
- Granite Switch 4.1 30B (GGUF Q4_K_M) · IBM
IBM's own llama.cpp GGUF build (Q4_K_M) of Granite Switch 4.1 30B.
- Granite Switch 4.1 30B (GGUF Q8_0) · IBM
IBM's own llama.cpp GGUF build (Q8_0) of Granite Switch 4.1 30B.
- LFM2.5-24B-A2B · Liquid AI
Liquid AI's MoE with 24B total and 2B active: fast on CPUs and edge devices.
- LFM2.5-24B-A2B (FP8) · Liquid AI
Liquid AI's own FP8 build (compressed-tensors) of LFM2.5-24B-A2B, 24B total, 2B active.
- Muse Glimmer 30B (GGUF Q4_K_M) · Meta
Meta's open Muse Glimmer 30B in Meta's own llama.cpp GGUF build (Q4_K_M).
- Muse Glimmer 30B (GGUF Q4_K_M + vision) · Meta
Meta's open Muse Glimmer 30B in Meta's own llama.cpp GGUF build (Q4_K_M); the projector reads images.
- ERNIE-Image-Turbo · Baidu
Baidu's 8B text-to-image DiT, distilled for few steps, with a Ministral prompt enhancer in the set.
- FrogNano 4B 2609 · Microsoft
Microsoft's compact repository-level coding agent, post-trained with RL from Qwen3.5-4B.
- FrogNano 4B 2609 (GGUF Q8_0) · Microsoft (GGUF: bartowski)
llama.cpp GGUF build (Q8_0, bartowski) of Microsoft's FrogNano coding agent, from Qwen3.5-4B.
- MiMo-V2.6-Distill-Qwen-9B · Xiaomi
Xiaomi's reasoning distillation of MiMo-V2.6 into Qwen3.5-9B.
- Mage-VL-Cowork (GGUF Q8_0 + vision) · Microsoft (GGUF: ggml-org)
llama.cpp GGUF build (Q8_0, ggml-org) of Microsoft's 4B computer-use model: reads screenshots, plans clicks and shell commands.
- Z-Image · Tongyi-MAI
Z-Image base, 6B: the full-step sibling of Z-Image Turbo; slower, for quality and fine-tuning.
- LongCat-Image · Meituan
Meituan's 6B text-to-image model with a Qwen2.5-VL text encoder; strong at Chinese and English text in images.
- MiMo-V2.6-Flash-RL (GGUF MXFP4 + vision) · Xiaomi (GGUF: ggml-org)
llama.cpp GGUF build (MXFP4, ggml-org) of Xiaomi's MiMo-V2.6-Flash-RL, 310B MoE; the projector reads images.
- Gemma 4 26B-A4B IT · Google
Gemma 4 MoE, 26B total and 4B active, instruction-tuned; chat with images.
- Gemma 4 E2B IT · Google
The smallest Gemma 4, instruction-tuned: 2B effective parameters for phones and laptops; chat with images.
- Mistral Large 4 1T-A52B (GGUF Q4_K_M) · Mistral AI (GGUF: ggml-org)
llama.cpp GGUF build (Q4_K_M, ggml-org) of Mistral Large 4: 1T MoE with 52B active. Commercial service use needs Mistral's licence.
- Mistral Large 4 1T-A52B (GGUF Q4_K_M + vision) · Mistral AI (GGUF: ggml-org)
llama.cpp GGUF build (Q4_K_M, ggml-org) of Mistral Large 4, 1T MoE with 52B active; the F16 projector reads images.
- gpt-oss-20b (GGUF MXFP4) · OpenAI (GGUF: ggml-org)
llama.cpp GGUF build (MXFP4, ggml-org) of OpenAI's gpt-oss-20b, a 21B MoE: its native 4-bit weights, kept packed.
- gpt-oss-120b (GGUF MXFP4) · OpenAI (GGUF: ggml-org)
llama.cpp GGUF build (MXFP4, ggml-org) of OpenAI's gpt-oss-120b, a 117B MoE: its native 4-bit weights, kept packed.
- GLM-5.3-Flash (GGUF UD-Q4_K_XL) · Z.ai (GGUF: unsloth)
llama.cpp GGUF build (UD-Q4_K_XL, Unsloth) of GLM-5.3-Flash, a 320B MoE with 18B active.
- GLM-5.3-Flash (GGUF UD-Q4_K_XL + vision) · Z.ai (GGUF: unsloth)
llama.cpp GGUF build (UD-Q4_K_XL, Unsloth) of GLM-5.3-Flash, a 320B MoE with 18B active. The BF16 vision projector reads images.
- GLM-5.3-Flash (GGUF Q8_0) · Z.ai (GGUF: unsloth)
llama.cpp GGUF build (Q8_0, Unsloth) of GLM-5.3-Flash, a 320B MoE with 18B active.
- LFM2.5-24B-A2B (GGUF Q4_K_M) · Liquid AI
Liquid AI's own llama.cpp GGUF build (Q4_K_M) of LFM2.5-24B-A2B, a MoE with 24B total and 2B active: fast on CPUs and edge devices.
- LFM2.5-24B-A2B (GGUF Q8_0) · Liquid AI
Liquid AI's own llama.cpp GGUF build (Q8_0) of LFM2.5-24B-A2B, a MoE with 24B total and 2B active: fast on CPUs and edge devices.
- LFM3 1.7B (GGUF Q8_0) · Liquid AI
Liquid AI's own llama.cpp GGUF build (Q8_0) of LFM3 1.7B: a small edge model for CPU, Mac and any GPU.
- LFM3-VL 2B (GGUF Q8_0 + vision) · Liquid AI
Liquid AI's own llama.cpp GGUF build (Q8_0) of LFM3-VL 2B: a small vision-language model for CPU, Mac and any GPU.
- MiMo-V2.6-Distill-Qwen-9B (GGUF Q8_0 + vision) · Xiaomi (GGUF: ggml-org)
llama.cpp GGUF build (Q8_0, ggml-org) of Xiaomi's reasoning distillation of MiMo-V2.6 into Qwen3.5-9B; the projector reads images.