You can deploy this model and start earning money today!

mkd-hossain/
Keural-Nova-v1.2-experimental-NVFP4

Name: Keural-Nova-v1.2-experimental-NVFP4
Availability: InStock

Views

No views yet

Keural Nova v1.2 — NVFP4 (experimental)

NVFP4 (4-bit) quantization of Keural Nova v1.2, by MKD, for high-throughput serving on NVIDIA Blackwell GPUs via vLLM.

67 GB (bf16) → 21.9 GB (NVFP4) — ~3× smaller.
Recipe replicates RedHatAI/Qwen3.6-35B-A3B-NVFP4: NVFP4 4-bit weights + activations on the language model; vision tower, lm_head, and MTP kept in high precision for accuracy.
Format: compressed-tensors / nvfp4-pack-quantized → served directly by vLLM.

⚠️ Hardware

FP4 tensor cores exist only on Blackwell (B200 / GB200 / RTX 50xx). Serve there for the speed/memory benefit. (Hopper/H200 lack FP4 kernels — this checkpoint is produced on H200 but runs accelerated on Blackwell.)

Serving (vLLM on Blackwell)

bash

vllm serve mkd-hossain/Keural-Nova-v1.2-experimental-NVFP4 \
  --served-model-name Keural-Nova-v1.2 \
  --tensor-parallel-size 2 \
  --max-model-len 262144 \
  --tool-call-parser qwen3_xml --enable-auto-tool-choice

Non-thinking default (the fixed chat template ships here); tool calls are Qwen XML.
Serve native 256K for agent/tool workloads (YaRN 1M only for genuine >256K requests).
Multimodal (vision) works — the vision tower is kept high-precision.

Quantization details

Tool: llm-compressor (compressed-tensors), NVFP4 scheme.
Calibration: 512 samples from the Keural Nova v1.2 SFT mix (Korean/English/code/tool-use), seq-len 2048.
Coverage: 31,070 linear layers quantized (incl. 30,880 MoE experts + attention + Gated-DeltaNet linear-attention); vision blocks + lm_head + MTP excluded (bf16).
Group scheme: weights 4-bit float / activations 4-bit float, per-block FP8 scales (NVFP4).

Accuracy

NVFP4 typically costs ~1–2% vs bf16 — the same small delta as the public nvidia/RedHatAI NVFP4 Qwen3.6-A3B checkpoints. Validate on your Blackwell against the bf16 base (Keural-Nova-v1.2-experimental) on your workload (tool-calling, Korean, code) before production.

License & attribution

Apache-2.0, inheriting the base Qwen/Qwen3.6-35B-A3B license. Quantization recipe after RedHatAI/Qwen3.6-35B-A3B-NVFP4. "Keural Nova" is a model by MKD.