DeepSeek-v4-Flash-IQ2

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

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.

Differences from the base checkpoint

Recorded in config.json under pulsar.deviations_vs_source:

• Fused scales. The base stores block-quantized weights as a 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).
• Six dead-weight tensors are absent: 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).

Loading

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.

Verification

Converted from the serving GGUF with the payload bytes preserved exactly, and checked per shard against that GGUF:

• structure — header parses, every offset inside the buffer, all offsets 32-byte aligned, no overlaps, file length exactly 8 + header_len + buffer_len;
• non-expert tensors — payloads byte-identical to the GGUF tensor ranges;
• routed experts — the 256 per-expert payloads of each projection, concatenated in order, byte-identical to the GGUF's stacked 3-D tensor;
• index closure — 36,909 index entries and 36,909 declared tensor names match in both directions, no orphans either way.

Provenance

• Base checkpoint: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp at revision 6821d6ad3681a4b137b066b76094fa82ebd0a380
• Converted from: v5-vexp-full256-iq2-t46.gguf (build-lane interchange, not a family artifact)
• Quantization: routed experts IQ2_XXS against the prisma Vision-Exp activation imatrix (ebb3d9da1745a8948e53063f8fb5fdf1f585dc9afbc577cf7ae8cfedb5333cf9, 129 entries); dense MXFP8; BF16 embedding/head; drafter experts MXFP4 (lossless from the base checkpoint's own FP4).