DFM-Mimir-v1.5-GGUF

DFM Mimir v1.5 — GGUF

Official GGUF conversions of DFM Mimir v1.5 from Danish Foundation Models. See the source model card for training, evaluation, intended uses and limitations. These files preserve the source checkpoint's weights, tokenizer and Gemma 4 chat template; quantization reduces weight precision for the Q8_0 and Q4_K_M variants.

Choose a file

Download one file. Each includes its tokenizer and chat template. Download size is not RAM use: inference needs additional memory, especially at longer contexts. Both quantized variants were produced directly from BF16, without requantizing an already quantized model.

hf download danish-foundation-models/DFM-Mimir-v1.5-GGUF dfm-mimir-v1.5-q4_k_m.gguf --local-dir ./mimir-v1.5

Runtime and chat template

Use the DFM Mimir app or patched llama.cpp with HRM-Text/PrefixLM support. Compatibility with arbitrary stock llama.cpp, Ollama and other GGUF runtimes is not implied. Always use the embedded Mimir chat template; the model was trained with that format, not raw untemplated prompts.

Context length is 4096 tokens. Longer contexts are experimental and need additional memory and validation. In the app, enable model discovery and select Check for newer models, or import a downloaded GGUF. Discovery lists official repositories containing both mimir and gguf in their names. Existing bundled models are not automatically replaced.

The tokenizer is loaded directly from tokenizer.json, as during training, without the export-time fix_mistral_regex rewrite. GGUF metadata uses tokenizer.ggml.pre=gemma4 and marks all 256 byte-fallback tokens. v1.5's tokenizer is byte-identical to the original Mimir tokenizer. Its Gemma 4 template adds support for mapping/object tool-response content; the supplied v1.5 template is preserved exactly.

Provenance and validation

• Source revision: cc57cebadf375947ced5ccd3317d9da6bf8f9677 of danish-foundation-models/DFM-Mimir-v1.5.
• Converter revision: 4122b9a814d5bd4f48f454367419f75c05ee5215 of schneiderkamplab/llama.cpp.
• Each precision passes 724/724 tokenizer/decoder/chat-template comparisons and 19/19 independent training-tokenizer audit cases.
• Each precision passes short Danish and English chats on CPU and Metal, using the embedded template, 1024-token context and a 32-token reply budget.
• Validation results are recorded in validation.json.
• Exact checksums: SHA256SUMS.
• Source/config/tokenizer hashes and conversion provenance: provenance.json.

These checks are bounded artifact qualification, not full quality benchmarks, long-context qualification, or new Linux/CUDA/Vulkan/mobile hardware testing. License: Apache 2.0, inherited from the source model.