---
title: "Together AI LoRA rank limit raised to 128, older SDKs now default to rank 128: 2026-09-29 | Harbyn"
description: "Together AI LoRA rank limit raised to 128, older SDKs now default to rank 128, announced Sep 29, 2026."
canonical: "https://harbyn.com/changes/together/2026-09-29-september-29-2026"
last-updated: 2026-10-02
---

# Together AI LoRA rank limit raised to 128, older SDKs now default to rank 128

Markdown version of <https://harbyn.com/changes/together/2026-09-29-september-29-2026>.

Behavior change - Together AI. medium severity. In effect.

Most fine-tuning models now accept a LoRA rank up to 128, up from 64; the default stays 64. Together CLI and Python SDK 2.36.0 use the default rank when lora\_r is not set, but earlier versions, including the 1.x SDK, use the model's maximum rank, now 128. Continuing training from a rank-64 adapter needs a matching rank, so upgrade to 2.36.0 or set lora\_r.

## Facts

| Fact | Value |
| --- | --- |
| Vendor | [Together AI](https://harbyn.com/vendors/together.md) |
| Kind | Behavior change |
| Severity | medium |
| Status | In effect |
| Announced | Sep 29, 2026 |
| Added to Harbyn | Oct 2, 2026 |
| Lead time | No dated deadline |
| Review | Reviewed by a person, extracted by a deterministic parser |

## Find it in your code

```sh
npx harbyn scan
```

This change names no identifier the scanner can match line by line, so the scan tells you whether your repository uses Together AI, from its packages and API hosts. Nothing leaves your machine.

## Sources

Where this event comes from: the page we read, when we read it, and the hash of that snapshot.

- <https://docs.together.ai/docs/changelog/rss.xml> (changelog, fetched 2026-10-02 01:24 UTC, sha256 `79d4503ce98977c3597fb0d0bc96b5ed878256eae3133b623a423e82821213f2`)

## Related changes

### More from Together AI

- [Batch API files retained for 7 days](https://harbyn.com/changes/together/2026-09-29-batch-api-files-retained-for-7-days.md) (Behavior change, announced Sep 29, 2026)
- [Qwen 27B fine-tuning batch size limits changed](https://harbyn.com/changes/together/2026-09-29-qwen-27b-fine-tuning-batch-size-limits-changed.md) (Behavior change, announced Sep 29, 2026)
- [lower serverless pricing for Qwen3.7-Max and Qwen3.8-Flash](https://harbyn.com/changes/together/2026-09-22-qwen-serverless-price-cut.md) (Pricing, effective Sep 22, 2026)
- [Together CLI v2.33.2 improvements](https://harbyn.com/changes/together/2026-09-10-preemptible-compute-for-gpu-clusters-preemptible-compute-is-now-in-public-pr-2.md) (Behavior change, announced Sep 10, 2026)

### Changes at other vendors

- [Amazon Selling Partner API: ORDER\_CHANGE notifications add PreFulfillmentStatusChange for Amazon Autos](https://harbyn.com/changes/amazon-sp-api/2026-10-14-sp-api-updates-new-listings-items-api-ivp-values-in-eu5-jp-and-ca-stores-oct-3.md) (effective Oct 28, 2026)
- [BigCommerce: 301 redirects will apply only to storefront channels](https://harbyn.com/changes/bigcommerce/2026-10-07-301-redirects-now-apply-to-storefront-channels-only.md) (effective Nov 10, 2026)
- [GitHub: Secret scanning audit log business field changes to enterprise slug](https://harbyn.com/changes/github/2026-10-20-updates-to-github-secret-scanning-7.md) (effective Nov 12, 2026)
- [Jira / Atlassian Cloud: New IP address ranges in use from January 1, 2027](https://harbyn.com/changes/atlassian-jira/2026-10-16-atlassian-cloud-adds-new-ip-address-ranges.md) (effective Jan 1, 2027)

[All upcoming deadlines](https://harbyn.com/deadlines.md)