Prism ML's ternary Bonsai 2 9B, based on Qwen3.5-9B, with thinking and tool calling. The Q2_0 packing requires upstream llama.cpp with Bonsai Hadamard support.
ReasoningVision
Models
Every model with a recorded recipe. Open a model for its Hugging Face card, hardware configurations, measured speeds and exact commands.
Prism ML's ternary Bonsai 2, built on Qwen3.6-27B: a 27B model packed small enough for 16 GB, with vision, tools and thinking.
ReasoningVision
DeepSeek's V4 Flash mixture-of-experts model (and its 0731 and DSpark builds).
Reasoning
DeepSeek's V4.1 Flash, a large mixture-of-experts model.
Reasoning
Google's discrete-diffusion Gemma 4 26B A4B: decodes blocks of tokens in parallel, with vision and thinking.
ReasoningVision
cyankiwi's AWQ INT4 quant of DiffusionGemma 26B A4B.
ReasoningVision
Google's Gemma 4 12B from May 2026: a compact multimodal model with thinking, good with images and documents.
ReasoningVision
Google's Gemma 4 mixture-of-experts model: 26B total, 4B active, so it runs fast while seeing images and thinking.
ReasoningVision
Google's dense Gemma 4, 31B parameters.
ReasoningVision
cyankiwi's AWQ INT4 quant of Gemma 4 12B.
ReasoningVision
cyankiwi's AWQ INT4 quant of Gemma 4 26B A4B.
ReasoningVision
cyankiwi's AWQ INT8 quant of Gemma 4 26B A4B.
ReasoningVision
cyankiwi's AWQ INT4 quant of Gemma 4 31B.
ReasoningVision
Z.ai's GLM-5.2 flagship, here as an NVFP4 + AQLM hybrid.
ReasoningVision
Z.ai's full GLM-5.3, the large sibling of GLM-5.3-Flash: frontier-class reasoning and agent work for a four-card workstation.
Reasoning
Z.ai's GLM-5.3-Flash, a large mixture-of-experts model for multi-GPU machines.
ReasoningVision
IBM's 30B dense model (64 layers). It thinks before it answers and calls tools.
Reasoning
Tencent's Hunyuan 3 (Hy3) 295B mixture-of-experts.
Reasoning
Thinking Machines' Inkling Small.
Reasoning
Poolside's Laguna S 2.1 coding model.
Reasoning
Mistral's Leanstral 1.5 for Lean 4 proof engineering.
Reasoning
Liquid AI's 2.6B hybrid model: tiny and fast, built for on-device work.
Text
Liquid AI's 24B hybrid mixture-of-experts (about 2B active; 30 short-convolution and 10 attention layers, 64 experts). It reasons before it answers and calls tools.
Reasoning
inclusionAI's Ling 3.0 Flash mixture-of-experts.
Reasoning
Xiaomi's MiMo V2.5 omni model.
ReasoningVision
Xiaomi's MiMo V2.6 Flash (RL build).
ReasoningVision
Meta's Muse Glimmer 30B.
Reasoning
Nanbeige4.3, a 33B dense Llama-architecture model with thinking, tool calls and a 512K window.
Reasoning
NVIDIA's hybrid Mamba-Transformer MoE (74B total, about 9B active), compressed from Nemotron 3.5 Super VL and quantization-aware distilled to a 24 GB checkpoint (IQ2_XXS experts, NVFP4 Mamba layers, Q8_0 attention), with an image projector.
ReasoningVision
NVIDIA's Nemotron 3.5 Lightning, a small fast mixture-of-experts.
Reasoning
NVIDIA's Nemotron Labs 3 Puzzle mixture-of-experts.
Reasoning
Ornith 1.5, the dense 9B.
Reasoning
Ornith 1.5, a 35B mixture-of-experts with built-in MTP.
Reasoning
Alibaba's 9B model from February 2026: small enough for an 8 GB card, with a 64K–128K window. It thinks, calls tools and stays quick on modest hardware.
ReasoningVision
Alibaba's dense Qwen3.6, 27B parameters.
ReasoningVision
Qwen3.6's 35B mixture-of-experts model with 3B active parameters.
ReasoningVision
Alibaba's 27B dense model from August 2026, and the strongest model that fits one consumer card. It thinks before it answers, calls tools reliably and reads images, which makes it the default for coding agents and long projects.
ReasoningVision
cyankiwi's AWQ INT4 quant of Qwen3.8-27B, linear attention kept in BF16.
ReasoningVision
cyankiwi's AWQ FP8 quant of Qwen3.8-27B.
ReasoningVision
bytkim's Qwen3.8-27B fine-tune for the Pi coding-agent harness, trained with supervised fine-tuning and reinforcement learning. Supports low, medium and xhigh reasoning; the EXL3 build retains a 4-bit MTP head.
ReasoningVision
Qwen3.8's large mixture-of-experts model: far more knowledge than the 27B, with few active parameters, so it stays fast.
ReasoningVision
StepFun's Step 3.7 Flash mixture-of-experts model.
Reasoning
StepFun's Step-5 preview, a ~600B mixture-of-experts model with image and video input.
ReasoningVision
UkisAI's post-training of Qwen3.8-27B for long-horizon agentic and coding work: it spends 58.5% fewer thinking tokens than Swift 1.0 while scoring slightly higher. The GSQ-RCO quants reuse ISTA-DASLab's published per-tensor allocation, refined with a Swift importance matrix, so IQ3_S lands at a KLD of 0.0417.
Reasoning
ConwayResearch's mixed 2-bit GGUF quantization of Qwen3.8-27B, using GSQ-RCO weights and an importance matrix for tool calling. This recipe serves text; the optional vision projector is not included.
Reasoning
Quaedra's 13.2 GB GGUF of Qwen3.6-35B-A3B: the routed experts quantized block-wise with GPTQ error feedback in standard llama.cpp formats, so the full 262k context fits on a 16 GB card.
Reasoning