MiniMax 2.5 is an open large language model targeting state of the art performance in coding, agentic tool use, search and office work, with reported scores of 80.2% in SWE-Bench Verified, 51.3% in Multi-SWE-Bench, and 76.3% in BrowseComp. The model has 230B parameters with 10B active, a 200K context window, and in unquantized bf16 format requires 457GB. Unsloth provides a Dynamic 3-bit GGUF quantization that reduces the size to 101GB (-62%) under the MiniMax-2.5 GGUF releases, using Unsloth Dynamic 2.0 to selectively upcast important layers to 8 or 16-bit for better quality, and supports multi GPU fine tuning workflows.
The usage guide centers on running MiniMax 2.5 locally through quantized GGUF files. The 3-bit dynamic quant UD-Q3_K_XL uses 101GB of disk space and is positioned to fit on a 128GB unified memory Mac for ~20+ tokens/s, and it also works faster with a 1x16GB GPU and 96GB of RAM for 25+ tokens/s. 2-bit quants or the biggest 2-bit configuration will fit on a 96GB device. For near full precision, the Q8_0 (8-bit) variant utilizes 243GB and is described as suitable for a 256GB RAM device or Mac for 10+ tokens/s. For best performance it is recommended that combined VRAM and RAM approximately equal the size of the chosen quantized model, although llama.cpp can offload to hard drive or SSD at the cost of slower inference.
MiniMax recommends default inference parameters of temperature = 1.0, top_p = 0.95, top_k = 40, repeat penalty = 1.0 or disabled, maximum context window of 196,608, and Min_P = 0.01, along with a default system prompt setting the assistant name to MiniMax-M2.5. The guide walks through running the 3-bit UD-Q3_K_XL quant using llama.cpp, starting with building llama.cpp from source with optional CUDA support, then loading the model directly via the llama-cli with a Hugging Face reference and a context size such as –ctx-size 16384, and tuning options like –threads, –n-gpu-layers and seed 3407. Users can download the quantized model weights with the huggingface_hub CLI using filters such as “*UD-Q3_K_XL*” or “*Q8_0*” to select between dynamic 3-bit and 8-bit variants. For production style deployments, the model can be served through llama-server with an alias and accessed via the OpenAI Python completion client pointing at a local HTTP endpoint, enabling chat completions compatible with existing OpenAI style tooling. A benchmark table compares MiniMax-M2.5 against MiniMax-M2.1, various Claude Opus, Gemini and GPT baselines across a wide range of reasoning, coding, search, tool use, office and domain specific tasks including AIME, GPQA-D, SciCode, IFBench, SWE-Bench variants, VIBE-Pro, BrowseComp, RISE, BFCL multi-turn, τ² Telecom, MEWC, GDPval-MM and Finance Modeling.
