GGUF header of the DSpark draft file (file name DeepSeek-V4-Flash-0731-DSpark.gguf, 10,896,057,568 bytes), read 2026-09-26
with a plain-Python reader over the key-value section only (5,248,323 bytes read; no tensor data touched).
magic b'GGUF', version 3, tensors 81, key-value pairs 62. Keys whose value is a long array or a string
of tokenizer data are summarised.

general.architecture = dflash
tokenizer.chat_template = {%- if not add_generation_prompt is defined -%}
  {%- set add_generation_prompt = false -%}
{%- endif -%}
{%- if not ...
general.type = model
general.sampling.top_p = 1.0
general.sampling.temp = 1.0
general.name = DeepSeek V4 Flash 0731
general.version = 0731
general.basename = DeepSeek-V4-Flash
general.size_label = 256x594M
general.license = mit
dflash.block_count = 3
dflash.context_length = 1048576
dflash.embedding_length = 4096
dflash.attention.head_count = 64
dflash.attention.head_count_kv = 1
dflash.rope.scaling.type = yarn
dflash.rope.scaling.factor = 16.0
dflash.rope.scaling.original_context_length = 65536
dflash.rope.scaling.yarn_beta_fast = 32.0
dflash.rope.scaling.yarn_beta_slow = 1.0
dflash.rope.freq_base = 10000.0
dflash.attention.layer_norm_rms_epsilon = 9.999999974752427e-07
dflash.expert_count = 256
dflash.expert_used_count = 6
dflash.expert_gating_func = 4
dflash.attention.key_length = 512
dflash.attention.value_length = 512
general.file_type = 38
dflash.rope.dimension_count = 64
dflash.attention.q_lora_rank = 1024
dflash.attention.sliding_window = 128
dflash.expert_feed_forward_length = 2048
dflash.expert_shared_count = 1
dflash.expert_weights_scale = 1.5
dflash.expert_weights_norm = True
dflash.swiglu_clamp_exp = [10.0, 10.0, 10.0]
dflash.swiglu_clamp_shexp = [10.0, 10.0, 10.0]
dflash.attention.indexer.head_count = 64
dflash.attention.indexer.key_length = 128
dflash.attention.indexer.top_k = 512
dflash.attention.output_group_count = 8
dflash.attention.output_lora_rank = 1024
dflash.attention.compress_ratios = [0, 0, 0]
dflash.attention.compress_rope_freq_base = 160000.0
dflash.hyper_connection.count = 4
dflash.hyper_connection.sinkhorn_iterations = 20
dflash.hyper_connection.epsilon = 9.999999974752427e-07
dflash.hash_layer_count = 0
dflash.block_size = 5
dflash.target_layers = [41, 42, 43]
general.quantization_version = 2
tokenizer.ggml.model = gpt2
tokenizer.ggml.pre = joyai-llm
tokenizer.ggml.tokens = [array of 129280 values]
tokenizer.ggml.token_type = [array of 129280 values]
tokenizer.ggml.merges = [array of 127741 values]
tokenizer.ggml.bos_token_id = 0
tokenizer.ggml.eos_token_id = 1
tokenizer.ggml.padding_token_id = 1
tokenizer.ggml.add_bos_token = False
tokenizer.ggml.add_eos_token = False
tokenizer.ggml.mask_token_id = 128799

has dflash.attention.sliding_window_pattern: False