This is the NVIDIA ModelOpt NVFP4 deployment checkpoint for migtissera/Tess-4-35B-A3B, pinned to source revision cae99edb934875a1977774502664dedf03c211db.

The model's routed MoE experts are packed as NVFP4, with calibrated FP8 KV-cache scales. The vision tower, multimodal projection path, shared experts, attention layers, embeddings, LM head, and included MTP head remain in BF16. The export retains the complete Qwen3.5 multimodal processor and the native 262,144-token context configuration.

Conversion details

GPU: one NVIDIA B200
NVIDIA ModelOpt commit: f479e7890f0d276e061d69b5a0d70d477ec863f2
Quantization format: nvfp4_experts_only
KV-cache format: calibrated FP8
Calibration data: 116 real Tess training conversations
Calibration window: one deterministic 4,096-token window per conversation
Export size: 25,615,467,408 bytes across three safetensor shards
MTP: all 19 source tensors preserved in BF16
Vision: all 333 source vision tensors preserved

vLLM

Use a recent vLLM build with ModelOpt NVFP4 and Qwen3.5 MoE VLM support:

bash

vllm serve migtissera/Tess-4-35B-A3B-NVFP4 \
  --quantization modelopt_fp4 \
  --kv-cache-dtype fp8 \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --trust-remote-code

The first validated deployment uses plain decoding. The preserved BF16 MTP head can be enabled later as a serving configuration change after validating the base NVFP4 path.