The 2-bit member of the family: routed experts stored as iq2_xxs_mmq_k
(2.0625 bpw). ~92 GB, so unlike the unqualified DeepSeek-v4-Flash (~168 GB)
this one fits a single GB10 at ~121 GB usable.
Derived from deepseek-ai/DeepSeek-V4-Flash-Vision-Exp.
This artifact IS a quantization, and the family name says so: the bare name means the source-precision repack, and a suffix marks a deviation from it. Routed experts here were quantized to IQ2_XXS with a measured activation imatrix; the unqualified build keeps the upstream checkpoint's own FP4.
| Layout | Used by | Bytes |
|---|---|---|
iq2_xxs_mmq_k |
main routed experts (43 layers × 256) | IQ2_XXS blocks, 66 B per 256 elements, k-major block order (2.0625 bpw) |
cutlass_mxfp4 |
drafter (MTP) routed experts (3 × 256) | per-expert E2M1 data then a swizzled E8M0 group-32 SF tile |
mxfp8_lt |
attention / dense / shared-expert projections | de-interleaved E4M3 then a swizzled E8M0 group-32 scale |
fp8_e4m3_soa_k |
one drafter head projection | E8M0 scale plane [rows][cols/32] then an E4M3 payload plane |
native |
norms, router, embeddings, hyper-connection params | plain BF16 / F32 / I32 as declared in the header |
Engine-layout tensors are stored as flat U8 blobs. Each shard's
__metadata__ is self-describing:
pulsar.tensors — per tensor: layout and dims_ne (the logical shape in
ggml ne[] order, the HF shape reversed).pulsar.experts — per routed-expert projection: n_experts, expert_bytes,
layout, contiguous, and the original stacked tensor name. Experts are
written back-to-back with no inter-expert padding, so a kernel addresses
them as base + xid * expert_bytes from one pointer.pulsar.kv_arch — the 53 architecture keys the engine reads; pulsar.kv
(primary shard only) carries the tokenizer block.format — pt; pulsar.alignment — 32.Recorded in config.json under pulsar.deviations_vs_source:
weight +
scale pair; 528 such .scale planes have no counterpart here because these
layouts carry the scales inside the same tensor.markov_head_w2 is stored in source order rather than reversed — the
quantizer transposes it deliberately.confidence_head.proj is rank-flattened to (4352,) from (1, 4352).layers.{0,1,2}.ffn.gate.bias
(hash-routed layers; the forward reads tid2eid for text and bias_vl for
image, never this bias) and mtp.{0,1,2}.ffn.gate.bias_vl (drafts are text).This runs on pulsar only. Stock transformers / vLLM cannot load it — they
have no decoder for the fused layouts, and U8 blobs have no meaning without
the metadata above. config.json's quantization_config.quant_method is
pulsar-safetensors so a standard loader fails loudly instead of misreading the
bytes.
Converted from the serving GGUF with the payload bytes preserved exactly, and checked per shard against that GGUF:
8 + header_len + buffer_len;deepseek-ai/DeepSeek-V4-Flash-Vision-Exp at revision
6821d6ad3681a4b137b066b76094fa82ebd0a380v5-vexp-full256-iq2-t46.gguf (build-lane interchange, not a
family artifact)ebb3d9da1745a8948e53063f8fb5fdf1f585dc9afbc577cf7ae8cfedb5333cf9,
129 entries); dense MXFP8; BF16 embedding/head; drafter experts MXFP4
(lossless from the base checkpoint's own FP4).model-00048-of-00048.safetensors 3.7 GBmodel-00046-of-00048.safetensors 3.6 GBmodel-00047-of-00048.safetensors 3.6 GBmodel-00045-of-00048.safetensors 2.1 GBmodel-00004-of-00048.safetensors 1.8 GBmodel-00012-of-00048.safetensors 1.8 GBmodel-00014-of-00048.safetensors 1.8 GBmodel-00016-of-00048.safetensors 1.8 GBmodel-00018-of-00048.safetensors 1.8 GBmodel-00020-of-00048.safetensors 1.8 GBmodel-00022-of-00048.safetensors 1.8 GBmodel-00024-of-00048.safetensors 1.8 GBmodel-00026-of-00048.safetensors 1.8 GBmodel-00028-of-00048.safetensors 1.8 GBmodel-00030-of-00048.safetensors 1.8 GBmodel-00032-of-00048.safetensors 1.8 GBmodel-00034-of-00048.safetensors 1.8 GBmodel-00036-of-00048.safetensors 1.8 GBmodel-00038-of-00048.safetensors 1.8 GBmodel-00040-of-00048.safetensors 1.8 GBmodel-00042-of-00048.safetensors 1.8 GBmodel-00044-of-00048.safetensors 1.8 GBmodel-00006-of-00048.safetensors 1.8 GBmodel-00008-of-00048.safetensors 1.8 GBmodel-00010-of-00048.safetensors 1.8 GBmodel-00013-of-00048.safetensors 1.8 GBmodel-00015-of-00048.safetensors 1.8 GBmodel-00017-of-00048.safetensors 1.8 GBmodel-00019-of-00048.safetensors 1.8 GBmodel-00021-of-00048.safetensors 1.8 GBmodel-00023-of-00048.safetensors 1.8 GBmodel-00025-of-00048.safetensors 1.8 GBmodel-00027-of-00048.safetensors 1.8 GBmodel-00029-of-00048.safetensors 1.8 GBmodel-00031-of-00048.safetensors 1.8 GBmodel-00033-of-00048.safetensors 1.8 GBmodel-00035-of-00048.safetensors 1.8 GBmodel-00037-of-00048.safetensors 1.8 GBmodel-00039-of-00048.safetensors 1.8 GBmodel-00041-of-00048.safetensors 1.8 GBmodel-00043-of-00048.safetensors 1.8 GBmodel-00005-of-00048.safetensors 1.8 GBmodel-00007-of-00048.safetensors 1.8 GBmodel-00009-of-00048.safetensors 1.8 GBmodel-00011-of-00048.safetensors 1.8 GBmodel-00002-of-00048.safetensors 1.8 GBmodel-00003-of-00048.safetensors 1.8 GBmodel-00001-of-00048.safetensors 941 MB有文件?提交种子。
提交种子