# Phase 4 — config metadata diff (GGUF vs HF) GGUF: `llama_cpp_models/mistralai_Mistral-Medium-3.5-128B-Q8_0/mistralai_Mistral-Medium-3.5-128B-Q8_0-00001-of-00004.gguf` HF: `mistral_medium_check/config.json` arch: `mistral3` | field | gguf_key | gguf_val | hf_path | hf_val | match | | --- | --- | --- | --- | --- | --- | | vocab_size | `mistral3.vocab_size` | `�` | `text_config/vocab_size` | `131072` | ✗ | | hidden_size | `mistral3.embedding_length` | `0` | `text_config/hidden_size` | `12288` | ✗ | | num_hidden_layers | `mistral3.block_count` | `X` | `text_config/num_hidden_layers` | `88` | ✗ | | num_attention_heads | `mistral3.attention.head_count` | ``` | `text_config/num_attention_heads` | `96` | ✗ | | num_key_value_heads | `mistral3.attention.head_count_kv` | `�` | `text_config/num_key_value_heads` | `8` | ✗ | | intermediate_size | `mistral3.feed_forward_length` | `p` | `text_config/intermediate_size` | `28672` | ✗ | | max_position_embeddings | `mistral3.context_length` | `�` | `text_config/max_position_embeddings` | `262144` | ✗ | | rope_theta | `mistral3.rope.freq_base` | `$tI` | `None` | `None` | — | | rms_norm_eps | `mistral3.attention.layer_norm_rms_epsilon` | `[np.float32(1e-05)]` | `text_config/rms_norm_eps` | `1e-05` | ✗ | | head_dim | `mistral3.attention.key_length` | `[np.uint32(128)]` | `text_config/head_dim` | `128` | ✗ | | sliding_window | `None` | `None` | `text_config/sliding_window` | `None` | — | | bos_token_id | `tokenizer.ggml.bos_token_id` | `�` | `text_config/bos_token_id` | `1` | ✗ | | eos_token_id | `tokenizer.ggml.eos_token_id` | `�` | `text_config/eos_token_id` | `2` | ✗ | | pad_token_id | `tokenizer.ggml.padding_token_id` | `�` | `text_config/pad_token_id` | `11` | ✗ | | rope_scaling.type | `mistral3.rope.scaling.type` | `yarn` | `None` | `None` | — | | rope_scaling.factor | `mistral3.rope.scaling.factor` | `[np.float32(64.0)]` | `None` | `None` | — | | rope_scaling.original_max_position_embeddings | `mistral3.rope.scaling.original_context_length` | `�` | `None` | `None` | — | | attn_logit_softcapping | `None` | `None` | `None` | `None` | — | | final_logit_softcapping | `None` | `None` | `None` | `None` | — | ## TL;DR **12 mismatch(es)** found: - `vocab_size`: gguf=`�` vs hf=`131072` - `hidden_size`: gguf=`0` vs hf=`12288` - `num_hidden_layers`: gguf=`X` vs hf=`88` - `num_attention_heads`: gguf=``` vs hf=`96` - `num_key_value_heads`: gguf=`�` vs hf=`8` - `intermediate_size`: gguf=`p` vs hf=`28672` - `max_position_embeddings`: gguf=`�` vs hf=`262144` - `rms_norm_eps`: gguf=`[np.float32(1e-05)]` vs hf=`1e-05` - `head_dim`: gguf=`[np.uint32(128)]` vs hf=`128` - `bos_token_id`: gguf=`�` vs hf=`1` - `eos_token_id`: gguf=`�` vs hf=`2` - `pad_token_id`: gguf=`�` vs hf=`11`