Sponsored by
Vmake Labs Video Enhancer
-O Vmake Labs Video Enhancer utiliza IA para melhorar a qualidade dos vídeos, restaurar detalhes, reduzir a desfocagem e aumentar
Model checkpoints for
pulsar
, a CUDA-only DeepSeek V4 inference engine for
the NVIDIA DGX Spark (GB10). Nothing here loads in stock
transformers
or
vLLM: the weights are stored in the byte layouts pulsar's kernels read, each
declared per tensor, and
quantization_config.quant_method
is
pulsar-safetensors
so a standard loader fails loudly.
How the family is named
<model>-<recipe>
The bare name is the faithful build.
DeepSeek-v4-Flash
is the
source-precision repack — every routed expert carries the upstream checkpoint's
own FP4 bytes, permuted into the layout the engine reads, with no quantizing.
A suffix marks a deviation from source precision:
suffix
meaning
(none)
source precision: routed experts repacked from the checkpoint's own FP4 (
cutlass_mxfp4
, 4.25 bpw)
-IQ2
routed experts quantized to IQ2_XXS (2.0625 bpw) to fit one Spark
-Mix-…
per-
(layer, parent)
promotion: some units at source precision, the rest at the lower rung
Dense projections are
mxfp8_lt
and the embedding/head BF16 in every variant;
only the routed experts move. That makes the family a one-dimensional spectrum
between the two flat endpoints, and a
-Mix-
artifact a point on it.
Model revision, population (full vs pruned) and provenance live in each model
card, not in the name.
cutlass_mxfp4
on promoted
(layer,parent)
units,
iq2_xxs_mmq_k
elsewhere
between
either
Sizes are the container total (48 shards + index). A single GB10 has ~121 GB
usable; the unqualified build deliberately needs the two-Spark rig.
Layouts
layout
bytes
cutlass_mxfp4
per-expert E2M1 data then a swizzled E8M0 group-32 scale-factor tile
iq2_xxs_mmq_k
IQ2_XXS 66 B per 256 elements, k-major block order
mxfp8_lt
de-interleaved E4M3 then a swizzled E8M0 group-32 scale
fp8_e4m3_soa_k
E8M0 scale plane then an E4M3 payload plane
native
plain
BF16
/
F32
/
I32
Every artifact is a 48-shard safetensors container in the upstream model's HF
shard scheme (
vision
,
layers.0..42
,
top
,
mtp.0..2
), with the engine's
architecture keys in each shard's
__metadata__
.
Verifying a download
Each model card lists its own gates. In general an artifact is expected to pass:
structure (gap-free, 32-byte aligned, exact length), expert contiguity
(
offset(E) == offset(0) + E*expert_bytes
), index closure in both directions,
and byte-identity of every payload against the artifact it was built from.
Runs of ElytronAI pulsar-artifacts on huggingface.co
0
Total runs
0
24-hour runs
0
3-day runs
0
7-day runs
0
30-day runs
More Information About pulsar-artifacts huggingface.co Model
pulsar-artifacts huggingface.co is an AI model on huggingface.co that provides pulsar-artifacts's model effect (), which can be used instantly with this ElytronAI pulsar-artifacts model. huggingface.co supports a free trial of the pulsar-artifacts model, and also provides paid use of the pulsar-artifacts. Support call pulsar-artifacts model through api, including Node.js, Python, http.
pulsar-artifacts huggingface.co is an online trial and call api platform, which integrates pulsar-artifacts's modeling effects, including api services, and provides a free online trial of pulsar-artifacts, you can try pulsar-artifacts online for free by clicking the link below.
ElytronAI pulsar-artifacts online free url in huggingface.co:
pulsar-artifacts is an open source model from GitHub that offers a free installation service, and any user can find pulsar-artifacts on GitHub to install. At the same time, huggingface.co provides the effect of pulsar-artifacts install, users can directly use pulsar-artifacts installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
