zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV

Qwen3.8-Flash-Next W4A16 + INT4-PLE + FP8-KV

All credit for the underlying quantization work belongs to halt95, author of the base Qwen3.8-Flash-Next-W4A16-Merlin checkpoint. This repository is that checkpoint repackaged with two additions: the 51B-parameter n-gram PLE embedding table converted from FP8 to INT4, and a calibrated FP8 KV cache. The W4A16 model body, the MTP head, and every other tensor are halt95's work, carried over unchanged. If this model is useful to you, go like and credit the original. For model capabilities and documentation see the base model card.

The INT4 PLE repack halves the pinned-host table from 47.7 GB to 26.8 GB; serving with VLLM_PLE_CPU_OFFLOAD=1 requires ~32 GB of free host RAM. The FP8 KV cache roughly doubles the token pool per GiB of VRAM and unlocks the model's native 262,144-token context.

To serve this model you must use a vLLM build that carries the int4 PLE plugin and the qwen4_exp model support. The zankich/vllm fork integrates both, and zankich/vllm-openai is the prebuilt serve image from that fork. Stock vLLM has neither the architecture nor the plugin.

Running the model

docker run -d --gpus all -p 8000:8000 \
  -v /path/to/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV:/model:ro \
  -e VLLM_PLE_CPU_OFFLOAD=1 \
  -e VLLM_QSA_KV_SCALES=/model/qsa_kv_scales_int4ple_2026-09-23_margin1.20.json \
  zankich/vllm-openai:latest \
  /model \
  --tensor-parallel-size 4 --enable-expert-parallel \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8_e4m3 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

• VLLM_PLE_CPU_OFFLOAD=1 is required: the 26.8 GB int4 PLE table is pinned in host RAM and looked up over UVA.
• --tensor-parallel-size 4 --enable-expert-parallel: the model serves across 4 GPUs with expert parallelism on the MoE layers.
• --quantization compressed-tensors: loads the W4A16 body.
• --kv-cache-dtype fp8_e4m3 with VLLM_QSA_KV_SCALES pointing at the sidecar: activates the calibrated FP8 KV cache. The sidecar carries per-layer static K/V scales (12 QSA layers, calibration margin 1.20).
• Vision is supported; the multimodal path serves alongside text.
• The reasoning/tool parsers make the model's think tags and tool calls parse correctly in chat completions; the chat template ships inside the checkpoint.

Optional, see the fork's documentation: --speculative-config for MTP speculative decoding, and the tiered KV offload connector (--kv-transfer-config with TieringOffloadingSpec).

Summary

Property	Value
Base model	halt95/Qwen3.8-Flash-Next-W4A16-Merlin (W4A16 GPTQ, compressed-tensors)
PLE n-gram table	INT4, group-32 symmetric, repacked from halt95's FP8 table
KV cache	FP8 E4M3, per-layer static scales in qsa_kv_scales_int4ple_2026-09-23_margin1.20.json (halt95's original calibration ships alongside as qsa_kv_scales_262k.json)
MTP head	bf16, unchanged
Size	95 GB total, 40 files
Host memory	~32 GB free RAM required for the pinned int4 table

INT4 PLE layout contract

text_config.ple_embedding_dtype: "int4". Words: int32 [rows, 20], 8 little-endian nibbles per word (element i of each 8-group in bits 4i). Codes stored bias+8: value = (nibble - 8) * scale. Scales: fp16 [rows, 5], one per group of 32 columns, with the FP8 source's global scale folded in (decode approximates fp8_source * global_scale). Round-trip residual bounded by half a quant step. ple-int4-manifest.json records per-shard SHA-256 digests and the packer's self-check residuals for the table repack.

Measured quality

MMLU-Pro (lm-eval, chat template, 5-shot, 700 questions): 0.8486 ± 0.0133. The published omlx 4-bit measurement of the base model is 0.843 on a 300-question subset. With the FP8 KV cache active: 0.875 ± 0.031 (112-question deterministic subset, calibration margin 1.20, clean tier) against 0.8839 for the same subset at the default KV dtype — parity within noise.