Hob-forge/BAR-5x7B-GGUF

BAR-5x7B — GGUF (first-of-its-kind FlexOlmo conversion)

This is the first GGUF conversion of `allenai/BAR-5x7B`, the largest member of AllenAI's BAR-family Mixture-of-Experts models released on 2026-04-19 based on the new FlexOlmo architecture.

5 experts × 7B → ~33B total parameters with top-k routing.

⚠ Requires patched llama.cpp

The FlexOlmo architecture is not yet supported in upstream `llama.cpp`. To run this GGUF use the FlexOlmo support fork:

• —Fork: https://github.com/Seraphiel102/llama.cpp/tree/flex-olmo-pr-clean

Build from the fork:

git clone https://github.com/Seraphiel102/llama.cpp.git
cd llama.cpp
git checkout flex-olmo-pr-clean
cmake -B build -DGGML_CUDA=OFF
cmake --build build -j --target llama-cli llama-quantize llama-completion

What FlexOlmo is

Per `transformers.models.flex_olmo`, FlexOlmoDecoderLayer is Olmo2's hybrid post-norm decoder layer with the dense FFN swapped for OlmoE-style top-k MoE routing. Specifically:

• —Attention with qnorm and knorm (Olmo2-style)
• —post_attention_layernorm and post_feedforward_layernorm (post-norm pattern, no input_layernorm)
• —Top-k MoE FFN with softmax routing (OlmoE-style)
• —No sliding-window attention

Files

Usage

./build/bin/llama-completion \
  -m BAR-5x7B.Q4_K_M.gguf \
  -p "The 5 experts in BAR-5x7B are " \
  -n 100

Validation

The Q4KM conversion was validated against the patched llama.cpp build using a basic arithmetic prompt and produces correct, coherent output.

Credit

• —Model: AllenAI — `allenai/BAR-5x7B`
• —FlexOlmo support in llama.cpp: PR by @Seraphiel102 / Nyx
• —Conversion: llama.cpp + the convert_hf_to_gguf.py patch from the support PR

If this saved you time, please ⭐ the llama.cpp PR.