Ollama v0.34.4-rc1 Released: XGrammar 0.2.7 Upgrade Improves Structured Output

Ollama Releases v0.34.4-rc1 Preview

Ollama has released v0.34.4-rc1, a release candidate focused on structured output improvements. The core change in this version is upgrading the XGrammar library in mlxrunner to version 0.2.7, bringing several fixes for structured output scenarios.

As one of the most popular local LLM runtime frameworks, Ollama has accumulated over 180,000 stars on GitHub with more than 18,000 forks, making it a key tool for locally deploying and running open-source large language models. This release candidate, tagged by developer jessegross, continues Ollama's ongoing effort to polish inference reliability.

What the XGrammar Upgrade Brings

XGrammar is a library focused on efficient and flexible structured generation, widely used to constrain LLM output formats and ensure models produce content conforming to predefined schemas (such as JSON Schema). In local inference scenarios, structured output capability directly determines whether a model can reliably generate data that downstream programs can parse.

According to the official release notes, upgrading to XGrammar 0.2.7 delivers two main categories of schema-related fixes:

Typed Dictionary Value Handling

The new version fixes issues with handling typed dictionary values. In practice, developers frequently need to constrain model outputs to dictionary structures with specific value types — for example, requiring a field to be a mapping from strings to integers. Previous versions could behave unexpectedly when processing these type-constrained dictionaries, and this update addresses those issues.

Improved Short Array Support

The second fix targets short array scenarios. The schema parsing and execution logic has been optimized for generating fixed-length or short-length array structures. This means that when an application requires the model to output short arrays — such as coordinate pairs or small lists of tags — the results will be more reliable.

XGrammar works by compiling format specifications like JSON Schema into finite state machines (FSMs) or context-free grammars (CFGs), then masking the candidate vocabulary at each token generation step to allow only tokens that conform to the current grammar state. This approach is known as "constrained decoding." Compared to post-hoc validation, it fundamentally eliminates the possibility of malformed output rather than relying on retry mechanisms. XGrammar's advantage over earlier similar libraries (such as Outlines and lm-format-enforcer) lies in its faster compilation of complex schemas and lower runtime inference overhead, enabling minimal performance impact in local inference scenarios. mlxrunner is Ollama's backend runtime for Apple Silicon (M-series) devices, so this upgrade means Mac users will directly benefit when using structured output.

Why Structured Output Matters

Structured output is a critical component for integrating LLMs into real production systems. When developers build AI agents, data extraction pipelines, or function-calling workflows, they rely heavily on models producing strictly formatted, parseable results. Any flaw in schema handling can cause downstream parsing failures or unexpected errors.

By continuously tracking updates to low-level constrained generation libraries like XGrammar, Ollama can quickly bring upstream community fixes into local runtime environments, giving users a more stable structured generation experience without needing to dig into underlying details. This kind of continuous iteration on the fundamentals is precisely what keeps open-source tools competitive.

What a Release Candidate Means

It's worth noting that v0.34.4-rc1 is a release candidate rather than a stable production release. These versions are typically used for broader testing before an official release, gathering community feedback and verifying that fixes are effective. Developers who want to try the structured output improvements early can test this release candidate; production environments that prioritize stability should wait for the subsequent official release.

While this update is modest in scope, it's a reliability improvement worth noting for developers who rely on Ollama for structured data generation. As local LLM use cases continue to expand, this kind of foundational polish will keep improving the overall user experience.

In the software release lifecycle, a Release Candidate (RC) comes after beta testing and before the general availability (GA) release. The naming convention (rc1, rc2, etc.) indicates which round of candidacy it represents — if a blocking issue is found during rc1, the team fixes it and ships rc2, continuing until no critical issues remain before promoting to a stable release. For active open-source projects like Ollama, RC versions are typically listed separately as "Pre-release" on the GitHub Releases page and are not automatically pushed to users via the standard ollama pull command — they require manually specifying the version number to install. This mechanism protects most everyday users from potential regressions while giving community contributors and early adopters a channel to participate in validation.