GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
wrldsuksgo2mars
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
wrldsuksgo2mars
Model Tree
deepseek-ai/DeepSeek-V4.1-Flash-1017
Input Modalities
Output Modalities
Supported Functionality
Byte-preserving storage repack of deepseek-ai/DeepSeek-V4.1-Flash-1017, revision 1bae0e1088993b9b82de08f59de5be5a34e42170, under the original MIT license (see LICENSE).
Only the two Engram embedding tables (layers 1 and 14) change storage. Each layers.{layer}.engram.embed.strided_{part:05} tensor is U8 [rows, 264]: bytes 0..256 are the original F8_E4M3 weight row, bytes 256..264 are the original F8_E8M0 scales. These bytes are opaque storage, not quantized integer values. All values are byte-identical; there is no requantization. Every row was compared and the deinterleaved planes matched full source tensor SHA256 hashes. Every untouched tensor retains its source bytes. Shards 1..46 are unchanged; 47/48 retain all other tensors unchanged.
config.json.engram_storage declares layout=fp8_row_interleaved_v1, weight_bytes=256, scale_bytes=8, tensor_suffix=strided, plus an ordered parts map per layer. Each part declares tensor and rows; cumulative row counts define its global range. Each table has three contiguous row-range parts in separate shards below 45 GB; concatenate their row spaces in declared order. Rewritten safetensors headers also carry engram_layout=fp8_row_interleaved_v1 in metadata. The original quantization_config is unchanged. A loader supporting this storage layout (CuteAFD) is required; unmodified Transformers does not load the interleaved table. GPU math is unchanged: the CPU gathers one contiguous row then splits existing weight/scale staging. No speed improvement is claimed. MAPPED_SCALE_PRELOAD_GIB does not preload interleaved scales; ENGRAM_WARM warms the whole table.
The original model implementation, tokenizer, license, and other repository assets are preserved. Refer to the source model card for model capabilities and usage requirements.