We value your privacy

We use cookies to enhance your browsing experience, serve personalized content and analyze our traffic. By clicking “Accept all” you consent to our use of cookies. Cookie Policy

More options
Skip to content
FacthoryFacthory
FacthoryFacthory
    • Workspace
    • Agent
    • Knowledge
    • Analytics
    • Office apps
    • Desktop
    All features
    What’s new
    Download
  • By work

    • Organizational Knowledge Risk & Gap Intelligence
    • Scientific Research & Experimentation
    • Autonomous Analytics & Insight Discovery
    • Workforce Readiness & Adaptive Learning
    • SOP Governance & Operational Standardization
    • Process Discovery & Continuous Optimization
    • All solutions

    By industry

    • Manufacturing
    • Automotive
    • Construction & Infrastructure
    • Real Estate
    • Public Sector
    • Life Sciences
    • All industries
  • Facthory for

    • Regulated industries
    • On-premises
    • Developers

    Discover

    • What’s new
    • Blog
    • Case studies
    • Webinars
    • Models

    Learn

    • Documentation
    • Help centre
  • Pricing
  • Download
Sign inGet started

A workspace for what your organization knows, with an agent that works on it.

Product

  • All features
  • Workspace
  • Agent
  • Knowledge
  • Analytics
  • Office apps
  • Desktop
  • Download
  • Pricing

Solutions

  • Organizational Knowledge Risk & Gap Intelligence
  • Scientific Research & Experimentation
  • Autonomous Analytics & Insight Discovery
  • Workforce Readiness & Adaptive Learning
  • SOP Governance & Operational Standardization
  • Process Discovery & Continuous Optimization
  • All solutions
  • All industries

Resources

  • Blog
  • Case studies
  • Webinars
  • Documentation
  • Help centre
  • Models
  • ROI calculator

Company

  • Who we are
  • Careers
  • Press hub
  • Brand kit
  • Philosophy
  • Contact us

Join our newsletter

Product news from Facthory, by email.

​

© 2026 Facthory. All rights reserved.
  • Privacy policy
  • Cookie policy
  • Terms and conditions
  • Imprint
  • Cookie settings
All modelsnvidia

Llama-3.2-1B-Qwen3.8-27B-Draft

text-generation
1.5B parameters
other
Scan not finished

The NVIDIA Llama-3.2-1B-Qwen3.8-27B-Draft model is an external draft model for Qwen3.8-27B, sharing the target's tokenizer and vocabulary. It proposes blocks of candidate tokens which Qwen3.8-27B verifies by rejection sampling, reducing decode latency.

View on Hugging FaceCompare with other models

Fact sheet

Parameters

1.5B

Context length

131,072 tokens

Architecture

LlamaForCausalLM

Shape

16 layers, hidden size 2,048, 32 attention heads, 8 key-value heads

Weights stored as

bfloat16

Task

text-generation

Languages

not published

Built on

meta-llama/Llama-3.2-1B, Qwen/Qwen3.8-27B

Released

11 Nov 2026

Last updated

17 Nov 2026

Downloads (last 30 days)

8

Licence and access

Licence

other (nvidia-open-model-agreement)

Access

Weights download without approval

Security

Hugging Face security scan

Scan not finished

Weight files by format

Safetensors: 1
Pickle-based: 0
GGUF: 0
Other formats: 0
Pickle-based files (.bin, .pt, .pth, .pkl, .ckpt) can run code when they are loaded; safetensors files cannot.

Facthory’s checks

Automated
Met
Safe weight formats

Every weight file is safetensors or GGUF.

Not evidenced
The hub’s malware and pickle scan

The scan has not finished.

Met
A declared licence

The card declares its licence.

Met
Published safety evaluations

The card reports safety or harm evaluations with their results.

Rated by Facthory’s crawler from the published facts; a person has not reviewed it yet.

GPU calculator

How much memory this model needs to serve at a quantization, context and number of concurrent sequences, which GPUs hold it, and how fast it can generate. Estimates from the model’s published size and shape, with the assumptions below.

Weights

BF16 / FP16 (16 bit)
​

Context per sequence: 8K tokens

Concurrent sequences: 1

KV cache

Memory

3.6 GB

Weights: 3 GB

KV cache: 0.3 GB

Runtime: 0.3 GB

GPUMemoryGPUs neededPer sequenceAll sequences
NVIDIA L424 GB
1 ×
56 tokens/s56 tokens/s
NVIDIA L40S48 GB
1 ×
160 tokens/s160 tokens/s
NVIDIA RTX 6000 Ada48 GB
1 ×
178 tokens/s178 tokens/s
NVIDIA RTX PRO 6000 Blackwell96 GB
1 ×
333 tokens/s333 tokens/s
NVIDIA A100 80GB SXM80 GB
1 ×
379 tokens/s379 tokens/s
NVIDIA H100 SXM80 GB
1 ×
622 tokens/s622 tokens/s
NVIDIA H200 SXM141 GB
1 ×
891 tokens/s891 tokens/s
AMD Instinct MI300X192 GB
1 ×
984 tokens/s984 tokens/s
AMD Instinct MI325X256 GB
1 ×
1,114 tokens/s1,114 tokens/s
Assumptions
  • Runtime memory (activations, buffers) is 10% of weights and KV cache, plus 1 GB per GPU for the framework.
  • A GPU is filled to at most 90% of its memory; tensor parallelism uses 1, 2, 4 or 8 GPUs in one server.
  • Generation is limited by memory bandwidth at 60% of the vendor’s peak: each step reads the weights once and every sequence’s KV cache. A mixture-of-experts model reads only its active experts, so it runs faster than shown.
  • GPU memory and bandwidth are the vendors’ published figures.

Profile

Kind

Base model

Lineage

Fine-tune

Family

Llama 3.2

Uses

Coding
Math
Reasoning
Conversation
Translation

In and out

Text → Text

Written on 17 Nov 2026 by Workers AI (@cf/zai-org/glm-4.7-flash) from the model card and the hub’s metadata: a reading of the publisher’s text, not checked by Facthory.

Published benchmarks

The model card publishes no benchmark results.

EU AI Act

What the EU AI Act asks of a general-purpose AI model depends on who places it on the EU market and how. These facts bear on it:

  • Licence: other
  • Access: Weights download without approval
  • Parameters: 1.5B
Automated
Rated by Facthory’s crawler from the published facts; a person has not reviewed it yet.

General-purpose AI model (Article 3(63))

Not determined. No estimate of its training compute is published, and the Commission’s indicative criterion (10²³ FLOP) needs one. If it is a general-purpose AI model, the obligations below bind its provider.

Systemic risk (Article 51)

Not presumed. No estimate of its training compute is published; without one, or a designation by the Commission under Article 52, Article 55 does not apply.

Open-source exemption (Article 53(2))

Does not apply. Its licence (other) is not a free and open-source licence, one that lets anyone use, study, change and share the model.

General-Purpose AI Code of Practice

NVIDIA is not on the European Commission’s list of signatories. The Code is voluntary.

What the card and the provider document

Partly met
Technical documentation published (Article 53(1)(a), Annex XI)

The card publishes some of the technical documentation’s items.

Not stated: training compute, energy used.
Partly met
Information for downstream providers (Article 53(1)(b), Annex XII)

The card gives some of the items downstream providers need.

Source: Read from Hugging Face on 17 Nov 2026, at commit a01c05f. huggingface.co/nvidia/Llama-3.2-1B-Qwen3.8-27B-Draft
How we rate
Not stated: intended uses.
Not evidenced
Copyright policy published (Article 53(1)(c))

We found no published copyright policy. The Act asks for one to be in place, not necessarily published.

Partly met
Public summary of training content (Article 53(1)(d))

The provider publishes training-content summaries; this model’s among them is not yet checked.

Source: developer.download.nvidia.com, read 8 Nov 2026
Met
Authorised representative in the EU (Article 54)

The provider names an authorised representative in the EU.

Source: developer.download.nvidia.com, read 8 Nov 2026
Not applicable
Adversarial testing (Article 55(1)(a))

Applies to models with systemic risk only.

Not applicable
Systemic risks assessed and mitigated (Article 55(1)(b))

Applies to models with systemic risk only.

Regulation (EU) 2024/1689, the EU AI Act · Source: General-Purpose AI Code of Practice, read 8 Nov 2026

Facthory’s EU AI Act score appears once a person has reviewed these checks. The articles we score against