---
title: "Effective Long-Context Scaling of Foundation Models"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-12; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2309-16039-effective-long-context-scaling.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Wenhan Xiong, Jingyu Liu, Igor Molybog, and 18 coauthors, Effective Long-Context Scaling of Foundation Models (CC BY 4.0).'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: "Not reproduced. Abs page and HTML key sections read 2026-11-12. Code not run."
category: arxiv
topic_area: 'ai-research: memory-and-context'
tags: arxiv, long-context, llama-2, continual-pretraining, position-encodings
arxiv_id: 2309.16039
alias: 2309-16039-effective-long-context-scaling
sources:
  - id: S1
    title: "Effective Long-Context Scaling of Foundation Models"
    publisher: arXiv
    authors: "Wenhan Xiong and 20 coauthors (listed on the arXiv abs page)"
    url: https://arxiv.org/abs/2309.16039
    version_read: "v3 (14 Nov 2023)"
    accessed: 2026-11-12
    license: 'CC BY 4.0'
---
# Effective Long-Context Scaling of Foundation Models

## Summary
The authors present long-context language models built through continual pretraining from Llama 2. The authors continually pretrain from Llama 2 with longer sequences and a dataset that upsamples long texts, producing models with effective context windows of up to 32,768 tokens. They report gains on regular tasks and larger gains on long-context tasks over Llama 2. With a cost-effective instruction-tuning step that does not need human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k on a suite of long-context tasks. Ablations find that abundant long texts in pretraining are not the key, and that long-context continual pretraining is more efficient than, and similarly effective to, pretraining from scratch with long sequences. [S1]

## Claims
- id: C1
  claim: "The models are built by continual pretraining from Llama 2 and support context windows of up to 32,768 tokens."
  source: S1
  evidence: "effective context windows of up to 32,768 tokens"
- id: C2
  claim: "The 70B instruction-tuned variant can surpass gpt-3.5-turbo-16k on a suite of long-context tasks."
  source: S1
  evidence: "the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance"
- id: C3
  claim: "The authors find continual pretraining with long sequences more efficient than, and similarly effective to, pretraining from scratch."
  source: S1
  evidence: "long context continual pretraining is more efficient and similarly effective"

## Verification
- Code available: No public code or weight URL is named on the abs page. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Base model: [2307-09288-llama-2](https://iinventedlight.com/library/notes/2307-09288-llama-2/). See also [2307-03172-lost-in-the-middle](https://iinventedlight.com/library/notes/2307-03172-lost-in-the-middle/), [2308-14508-longbench](https://iinventedlight.com/library/notes/2308-14508-longbench/), and the positional-interpolation note in this batch.
- Limits: The 32,768-token figure and the gpt-3.5-turbo-16k comparison are as of the v3 paper.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 27 Sep 2023 (v1); last revised v3 (14 Nov 2023). Sources: [S1] https://arxiv.org/abs/2309.16039, v3 (14 Nov 2023). checked_on: 2026-11-12.

## When to apply
Use when asking how Meta extended Llama 2-style models to 32k context, or how continual pretraining compares with training from scratch.

## Facts
- id: F1
  claim: "The work extends Llama 2 by continual pretraining to context windows of up to 32,768 tokens."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: "They report that the 70B instruction-tuned variant can surpass gpt-3.5-turbo-16k on their long-context suite."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: "Ablations find continual pretraining more efficient than training from scratch with long sequences."
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2309.16039, https://arxiv.org/abs/2309.16039. License: CC BY 4.0. v3 (14 Nov 2023). Accessed 2026-11-12.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
