CYBER-FROST-3.8 GGUF

The queued trunk quants are uploaded. Each one passed a short Vulkan generation test. CYBER-FROST-3.8-Q2_K_S.gguf now carries the MTP head inside the file. The separate mtp- files still do not load.

Source checkpoint: Blackfrost-AI/CYBER-FROST-3.8-BF16. Architecture qwen4exp (Qwen4ExpForConditionalGeneration), from Qwen/Qwen3.8-Flash-Next. The vision tower is in mmproj-CYBER-FROST-3.8-f16.gguf. One mmproj works with every trunk.

About 180B parameters, about 6B active. 48 layers, 512 experts, 10 experts per token, one shared expert. Configured context is 262144. Smoke tests used 512 tokens of context. Long context is not tested.

The checkpoint includes one native MTP draft block, trained before the final trunk behavior pass. That block is grafted into CYBER-FROST-3.8-Q2_K_S.gguf as layer 48, Q8_0, from mtp-CYBER-FROST-3.8-Q8_0.gguf. The other trunks do not have it yet. The standalone mtp- files still do not load.

Files

Trunk files are the model. mtp- files are only a draft head. You do not chat with an mtp- file by itself.

Quant choices

No importance matrix. These dynamic files are a per-layer type mix. They are not Unsloth Dynamic 3.0, which also depends on their calibration set.

Q4_K_M is one type for most 2D weights. output.weight and token_embd.weight are Q6_K. Attention v is Q5_K.

UD-Q4_K_XL keeps the first and last layers, and attention and SSM tensors, at Q5_K or Q6_K. Output and token embeddings are Q8_0. The GGUF file-type field still says Q4_K_M.

UD-IQ4_XS uses that same layer split with IQ4_XS where the row width allows it, and IQ4_NL where it does not. No imatrix, so this will not match a calibrated IQ4_XS.

MXFP4_MOE puts MXFP4 on expert tensors and the n-gram table only. Attention, SSM, output, and embeddings stay Q8_0.

On the 3-bit and 2-bit files, two tensors cannot use a 256-wide block, so they are Q4_0:

• Expert down projections, row length 640.
• The n-gram table per_layer_token_embd, row length 160.

On the 4-bit files those two tensors are Q5_0 (flat Q4_K_M) or Q5_1 (dynamic files).

ssm_conv1d.weight and ple_conv1d.weight are stored as F32. Files uploaded before 2026-10-08 had them as F16. Stock llama.cpp then crashed with GGML_ASSERT(src1->nb[0] == sizeof(float)) when any layer ran on CPU, for example -ngl 0 or partial offload. Only those 37 small tensors changed. The rest of each file is byte-identical. Thanks to Nico for the report.

Norms and the router stay float32.

Run a trunk

Tested on llama.cpp b1-4da6337, Radeon 680M, Vulkan, with async uploads disabled. The file is mmap'd. Experts and the n-gram table stay on CPU and fault in from disk. GTT stayed near 45 MB. Fit is off so the loader does not try to fill the GPU with the expert pool.

KV cache is f16. Quantized KV crashes this architecture.

Smoke context was 512. Prompt speed was about 0.5 to 0.7 tok/s. Generation was about 1.0 to 1.1 tok/s.

GGML_VK_DISABLE_ASYNC=1 llama-cli \
  -m CYBER-FROST-3.8-Q4_K_M.gguf \
  -ot "per_layer_token_embd=CPU,exps=CPU" \
  -cmoe \
  -ngl 99 \
  -c 512 \
  -lm mmap -fit off \
  -ctk f16 -ctv f16 \
  --jinja

UD-Q4_K_XL uses the same flags.

Draft head

CYBER-FROST-3.8-Q2_K_S.gguf is the only trunk with the head in the file. Run it with --spec-type draft-mtp. Do not pass -md.

GGML_VK_DISABLE_ASYNC=1 llama-cli \
  -m CYBER-FROST-3.8-Q2_K_S.gguf \
  -ot "per_layer_token_embd=CPU,exps=CPU" \
  -cmoe \
  -ngl 99 \
  -c 512 \
  -lm mmap -fit off \
  -ctk f16 -ctv f16 \
  --spec-type draft-mtp --spec-draft-n-max 2

This needs the qwen4exp MTP graph in the local llama.cpp build on this machine. The head reads the wide residual from before the final mixer, not a separate draft file.

mtp-CYBER-FROST-3.8-Q8_0.gguf still does not load next to a trunk. llama.cpp stops because output_hc_norm.weight is missing. The mixer weights are stored as blk.48.nextn.hc_head_norm.weight. The Q4_K_M draft was exported the same way. Do not pass either sidecar with -md.

Vision

mmproj-CYBER-FROST-3.8-f16.gguf is the stock Qwen3-VL ViT (qwen3vl_merger): 27 blocks, patch 16, spatial merge 2, no deepstack. It projects to 2560, which matches every trunk. Converted with convert_hf_to_gguf.py --mmproj --outtype f16.

On Vulkan, pass --no-mmproj-offload. The vision encoder on Vulkan stops at GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size()). On CPU it encodes a 256x256 image in about 3 s.

GGML_VK_DISABLE_ASYNC=1 llama-mtmd-cli \
  -m CYBER-FROST-3.8-Q2_K_S.gguf \
  --mmproj mmproj-CYBER-FROST-3.8-f16.gguf \
  --no-mmproj-offload \
  --image image.png \
  -p "Describe this image." \
  -ot "per_layer_token_embd=CPU,exps=CPU" \
  -cmoe \
  -ngl 99 \
  -c 1024 \
  -lm mmap -fit off \
  -ctk f16 -ctv f16

A 256x256 image is 64 tokens. Prefill of those tokens took about 59 s. Only Q2_K_S is tested with images. The other trunks use the same graph and metadata.

Test log

Prompt: write a Python 3 function named add that takes two ints and returns their sum. Pass means the reply contains that function and a return.

Image test: 256x256 PNG, red square and blue circle on yellow, the word "CAT" in black. Prompt: describe the shapes, colors, background, and any text.

License

Qwen Community License 1.0, same terms as the source checkpoint. See LICENSE.