DEV.co
Open-Source LLM · sakamakismile

Qwen3.6-27B-Text-NVFP4-MTP

Qwen3.6-27B-Text-NVFP4-MTP is a quantized text-only variant of Alibaba's Qwen3.6-27B base model. It uses NVIDIA's NVFP4 quantization format (modelopt native) and restores a Multi-Token Prediction (MTP) head in bf16 to enable speculative decoding. Designed for vLLM on NVIDIA Blackwell (SM120) hardware, it achieves ~1.7× throughput improvement over the baseline compressed-tensors variant on concurrent long-form decodes. The model is 27.78B parameters, supports 256K context length (trained max), and handles text generation in 12+ languages.

Source: HuggingFace — huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP
16.7B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
339.8k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developersakamakismile
Parameters16.7B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads339.8k
Likes80
Last updated2026-04-29
Sourcesakamakismile/Qwen3.6-27B-Text-NVFP4-MTP

What Qwen3.6-27B-Text-NVFP4-MTP is

NVFP4-quantized 27.78B-parameter causal language model based on Qwen3.6-27B. Quantization uses nvidia-modelopt 0.43.0 with NVFP4_DEFAULT_CFG (calibrated on 20 samples from neuralmagic/calibration, max_seq_len 8192). Vision tower stripped; MTP head (1 layer, ~850 MB bf16) and linear-attention convolutions (48/64 layers) kept in bf16; lm_head kept in bf16. Model footprint ~14 GB after quantization + bf16 components. Supports speculative decoding with num_speculative_tokens=3 via single MTP layer applied recursively (mean acceptance length ~3.0–4.0 at n=3). Verified on RTX PRO 6000 Blackwell with vLLM 0.19.1rc1.

Quickstart

Run Qwen3.6-27B-Text-NVFP4-MTP locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-throughput inference on Blackwell hardware

Optimized for vLLM on RTX PRO 6000 / RTX 5090 Blackwell. Achieves 189–207 tok/s aggregate throughput on 2-parallel 256K-context decodes (vs 119 baseline), making it suitable for concurrent multi-request serving where latency and throughput are both critical.

Long-context document summarization and retrieval augmentation

256K context window enables processing of multi-document inputs, long essays, and code repositories in a single forward pass. Useful for RAG systems, legal/medical document analysis, and repository-scale code generation.

Cost-effective self-hosted inference

Quantized model (~14 GB) fits in consumer/workstation Blackwell cards (RTX 5090, RTX PRO 6000) without enterprise GPU costs. Apache 2.0 license permits unrestricted deployment in private clouds or on-premises.

Running & fine-tuning it

ESTIMATE: ~14–15 GB VRAM (NVFP4-quantized model ~13 GB + bf16 MTP/conv1d/lm_head ~850 MB + vLLM overhead). Tested on NVIDIA RTX PRO 6000 Blackwell (SM120). Also expected to work on RTX 5090 and other Blackwell cards with ≥16 GB VRAM. Requires `--gpu-memory-utilization 0.85–0.9` to fit 256K context with KV cache in FP8. Older architectures (A100, H100, RTX 4090) are not verified; vLLM's modelopt native path is Blackwell-specific.

No guidance provided in the model card. Quantized models (NVFP4) are typically not suitable for full fine-tuning due to frozen weights; QLoRA or prompt-tuning on the base bf16 Qwen3.6-27B is an alternative. Contact the maintainer (sakamakismile) or defer to the base Qwen3.6-27B documentation for tuning strategies. MTP head is restored in bf16, but its fine-tuning feasibility is Unknown.

When to avoid it — and what to weigh

  • Vision/multimodal input required — Vision tower is physically deleted in this variant to save memory and reduce model size. If you need image-to-text or VLM capabilities, use the full Qwen3.6-27B or the VLM-capable Qwen3.6-27B-NVFP4 (non-text-only) instead.
  • Non-Blackwell GPU or low-VRAM hardware — Model is specifically optimized for Blackwell SM120 (vLLM native modelopt path). Performance on older architectures (A100, H100, 4090) unverified. Requires ~15 GB VRAM for inference; not suitable for sub-16GB cards.
  • Guaranteed exact-match output reproducibility or determinism requirements — Quantization (NVFP4) and speculative decoding introduce numerical divergence vs the fp32/bf16 baseline. For safety-critical or audit-heavy use cases, benchmarking against your tolerance is mandatory.
  • Very small batch sizes or single-token inference latency critical — Speculative decoding shine on batch/long-form decodes; single short-request latency (50-token prompt) is ~14 ms (1/71 tok/s) without speculation, ~7.5 ms (1/132 tok/s) with n=3 speculation. Prefill is not heavily optimized in the card.

License & commercial use

Apache 2.0 (inherited from base Qwen3.6-27B). Permissive OSI-approved license permitting commercial use, modification, and redistribution under attribution.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use, derivative works, and private deployment. No proprietary restrictions, gating, or commercial licensing fees apply. You may use this model in commercial products (SaaS, closed-source, for-profit) without seeking permission from Alibaba or sakamakismile, subject only to retaining the Apache 2.0 notice. However, verify integration with any proprietary vLLM or NVIDIA software licenses if used in a commercial service.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Quantized model artifacts are safetensors format (no arbitrary code execution on load if using trusted transformers library). No security audit or adversarial robustness claims are made. MTP head is restored in bf16, but integrity depends on the open-source quantization pipeline (lna-lab/GGUF-to-NVFP4-SM120) being properly reviewed. Recommend standard LLM safety practices: input sanitization, output filtering, rate limiting, and monitoring for misuse in production. Quantization may reduce model's ability to recognize certain adversarial or harmful patterns; no impact assessment provided.

Alternatives to consider

Qwen3.6-27B-NVFP4 (compressed-tensors variant)

Same base model but uses slower compressed-tensors quantization format and MTP head is dropped (0% draft acceptance). ~1.7× slower on concurrent long-form decodes than this -Text-NVFP4-MTP variant. Choose if vision/VLM capability is needed (this one is text-only).

Osoleve/Qwen3.5-27B-Text-NVFP4-MTP

Qwen3.5 predecessor (not 3.6); same quantization and MTP-restoration approach. Smaller context window (4K trained max vs 256K). Use if you need the Qwen3.5 architecture or if 3.6 is not available in your region/tool.

Official Qwen/Qwen3.6-27B (bf16 base model)

Full-precision baseline (~55 GB VRAM required). No quantization loss; enables fine-tuning and full capability access. Choose if you have ample GPU memory, need exact output reproducibility, or plan to customize the model extensively.

Software development agency

Ship Qwen3.6-27B-Text-NVFP4-MTP with senior software developers

Qwen3.6-27B-Text-NVFP4-MTP delivers 1.7× faster inference on Blackwell hardware via optimized quantization and speculative decoding. Perfect for high-volume, long-context inference with Apache 2.0 commercial freedom. Start a proof-of-concept with vLLM or contact our AI infrastructure team to integrate into your self-hosted or cloud deployment.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwen3.6-27B-Text-NVFP4-MTP FAQ

Can I use this model commercially in a SaaS product?
Yes. Apache 2.0 is a permissive OSI license that allows commercial use without restriction. You may deploy it in a for-profit SaaS, sell inference services, or embed it in proprietary software, provided you retain the Apache 2.0 license notice. Verify compatibility with any proprietary vLLM or NVIDIA software you bundle.
What GPU do I need to run this?
Minimum NVIDIA Blackwell (SM120): RTX PRO 6000, RTX 5090, or equivalent with ≥16 GB VRAM. Verified on RTX PRO 6000 Blackwell. Older architectures (A100, H100, RTX 4090) are not officially supported; vLLM's native modelopt path is Blackwell-specific. Model footprint is ~14–15 GB including KV cache and vLLM overhead at 256K context.
Why is speculative decoding (MTP) better than vanilla decoding?
The restored MTP head allows vLLM to draft multiple tokens per step using a single extra forward pass. At num_speculative_tokens=3, mean acceptance length is ~3–4 tokens, yielding 1.5–2× per-token decode speedup with no latency penalty. Baseline Qwen3.6-27B-NVFP4 drops the MTP head during quantization, so it cannot use this optimization.
Can I fine-tune this quantized model?
Not easily. NVFP4 quantization freezes most weights; full fine-tuning is not practical. QLoRA or prompt-tuning on the base bf16 Qwen3.6-27B is an alternative. No guidance is provided in the card; contact sakamakismile or refer to the base Qwen documentation for tuning strategies.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Qwen3.6-27B-Text-NVFP4-MTP is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy a High-Throughput Quantized LLM?

Qwen3.6-27B-Text-NVFP4-MTP delivers 1.7× faster inference on Blackwell hardware via optimized quantization and speculative decoding. Perfect for high-volume, long-context inference with Apache 2.0 commercial freedom. Start a proof-of-concept with vLLM or contact our AI infrastructure team to integrate into your self-hosted or cloud deployment.