GaelicThunder/DeepSeek-V4-Flash-Vision-Exp-ablit-EXL3-Kalibrated

Verified creator GaelicThunder verified
🤗 Hugging Face sourceimage-text-to-textmit116 GBsafetensorsChecksums witnessedupdated today
No torrent yet

Kalibrated Vision Exp — DeepSeek-V4-Flash-Vision-Exp (abliterated), EXL3, one DGX Spark

Kalibrated Vision Exp is vcruz305's MixedK pack of DeepSeek-V4-Flash-Vision-Exp with 22 more expert layers re-quantized at a calibrated 3-bit. The K in the name is MixedK's; the "calibrated" is what was added. It is ready to serve on one NVIDIA DGX Spark (GB10, 128 GB) with the recipe in GaelicThunder/DeepSeek-V4-Flash-Vision-One-DGX-Spark: text + images, 245,760 tokens of context, speculative decoding, CUDA graphs.

git clone https://github.com/GaelicThunder/DeepSeek-V4-Flash-Vision-One-DGX-Spark
cd DeepSeek-V4-Flash-Vision-One-DGX-Spark
./start.sh        # downloads this repo (106 GB), serves it, waits for /v1/models, prints a test request

What is in the pack

vcruz305's MixedK Kalibrated Vision Exp
base model DeepSeek-V4-Flash-Vision-Exp, abliterated by drowzeys (wo_b of layers 10–35, λ 3.5) same weights, same abliteration
routed experts per layer 256 256 (none pruned)
expert bits 2-bit trellis, 6 layers at 3-bit 3-bit on 28 layers (his 6 + 22 calibrated), 2-bit on 15
everything else BF16 non-expert tensors his tensors, requantized to the FP8 block format the sparkinfer image loads (the same conversion the MixedK route of the recipe applies)
layout 48 exllamav3 shards rank-sliced tp1 (tp1/) + the DSpark K64 draft (dspark-draft-k64/), i.e. exactly what scripts/serve.sh mounts
size 95 GB 106 GB; 100.08 GiB resident

The 22 added layers are 27 23 31 35 32 34 37 40 1 39 25 29 8 30 19 26 24 11 12 9 0 42 (ranked by the measured quantization error of their expert tensors); MixedK's own 3-bit layers 3 13 21 22 28 41 are kept as they are. The other 15 expert layers, attention, indexer, shared experts, router rows, hash tables, head, vision tower and the draft's expert plan are vcruz305's files, unchanged. Same EXL3 trellis format, same mcg codebook.

Quality

Paired per-token NLL on 64,859 frozen tokens (wikitext · gsm8k · code), greedy, the same items for every pack, against the original-precision release of the same abliterated model (FP8 attention, MXFP4 experts, served with vLLM on a 2×H200 pod). "% kept" is exp(−Δ mean NLL): the geometric mean over tokens of p_pack / p_original. 100 % is the original.

% of the original's token probability kept prose math code all
Kalibrated Vision Exp 87 % 99.8 % 92 % 90 %
MixedK (vcruz305) 76 % 94 % 85 % 81 %
REAP-216 3-bit of the same model (216 experts, all 3-bit) 57 % 98 % 94 % 70 %

Kalibrated − MixedK, paired on the same tokens: −0.133 ± 0.005 nats on prose, −0.062 ± 0.008 on math, −0.081 ± 0.005 on code. MMLU-Pro (251 items, two option orders): 63.2 % vs 64.3 %, a tie within the harness noise of ±1 item. Perplexity on 8 fixed passages 4.487 vs 4.545. The 0731 text packs (0xSero / MiaAI-Lab) are a different base model (V4-Flash-0731) and are not in this table; the cross-model comparison with its caveat, the experts-versus-bits study with byte-identical weights, the layer choice and every receipt are in docs/KALIBRATED.md.

Serving on one GB10

measured
context 245,760 tokens per request; KV pool 269,471 tokens at gpu_memory_utilization 0.925
decode, DSpark K5 draft 37.6 tok/s code · 34.1 counting · 19.7 free prose with thinking (medians of 3)
load 28 s with a warm page cache; the first boot compiles ~10 min of kernels
images 1 to 3 per prompt verified, same endpoint (image_url)

The memory budget: with the non-expert tensors in FP8 and CUDA graphs (the sparkinfer image) about 16 GiB more are free than on the PyPI vLLM + plugin stack, and this pack spends them on expert bits rather than on a bigger KV pool (MixedK on the same image holds 986k tokens). 26 promoted layers load but leave no room for one 245k request, so the pack stops at 22. Never set the utilization above 0.93 on a 128 GB Spark.

Known issue, mitigated: the serving stack occasionally emits the token )Skip at clause ends (a decode-path logits fault, documented in the recipe). start.sh puts a small proxy in front of the engine that adds "bad_words": [")Skip", ",Skip", ".Skip"] to every request; if you serve the pack another way, add it yourself.

How it was built

exllamav3 (0531096 + the recipe's conversion patch) on a 2×H200 pod, a per-tensor recipe with 3-bit on the chosen layers, -cb mcg -cr 250 -cpi 600, the non-promoted layers stored at 16 bit during the run so the calibration activations stay exact; each finished layer spliced into a 256-expert file in MixedK's exact 3-bit layer layout. Scripts: scripts/kalibrated/.

Credits

  • vcruz305 — this is his MixedK pack with 22 layers added, nothing less: the 2-bit experts, the non-expert tensors, the vision tower, the six original 3-bit layers and the draft plan are his files, and his recipe found the same load_weights bug on the same day. If he wants the 22 layer files folded into the MixedK pack itself, they are his to take: same format, same layout, same license.
  • DeepSeek for DeepSeek-V4-Flash-Vision-Exp; drowzeys for the abliteration the weights carry.
  • 0xSero for the sparkinfer image and the rank-sliced tp1 layout; MiaAI-Lab for the single-Spark recipe; turboderp for EXL3 and a converter that takes a per-tensor recipe.

License and terms

DeepSeek-V4-Flash-Vision-Exp is released under the MIT license; this pack carries the same license. The weights have had their refusals removed upstream (see the source's terms): they are published for red-teaming, security research and evaluation, and you are responsible for your deployment and jurisdiction.