Ollama v0.21.1, released on , introduces the Kimi CLI for agentic execution tasks and significantly enhances MLX runner performance. This update improves sampling speed, adds logprobs support, and refines prompt tokenization and thread safety for MLX-compatible models, alongside other fixes.
| Fact | Detail |
|---|---|
| Released by | Ollama |
| Release date | |
| What it is | An update to the Ollama platform, introducing Kimi CLI and various performance enhancements. |
| Who it is for | Developers and users working with local AI models, especially those using MLX or agentic systems. |
| Where to get it | Not yet disclosed. |
| Price | Not yet disclosed. |
- Ollama v0.21.1 introduces the Kimi CLI for agentic execution.
- MLX runner now supports logprobs for compatible models.
- MLX sampling is faster due to fused top-P and top-K.
- MLX prompt tokenization is improved via request handler goroutines.
- GLM4 MoE Lite performance is enhanced with a fused sigmoid router head.
- The Kimi CLI, accessible through Ollama, enables multi-agent systems for complex tasks [Source event].
- MLX runner in v0.21.1 now includes logprobs support for compatible models [Source event].
- MLX sampling speed is improved by fusing top-P and top-K in one pass [Source event].
- Prompt tokenization for MLX is enhanced by moving it into request handler goroutines [Source event].
- GLM4 MoE Lite sees performance gains from a fused sigmoid router head [Source event].
- Structured outputs for Gemma 4 models are fixed when
think=false[Source event]. - The model picker in the macOS app no longer shows stale models after chat switching [Source event].
What is Ollama v0.21.1
Ollama v0.21.1 is an update to the Ollama platform, released on , that introduces the Kimi CLI and various performance and stability improvements [Source event]. This version focuses on enhancing the MLX runner, improving agentic execution capabilities, and fixing specific bugs [Source event].
What is new vs the previous version
Ollama v0.21.1 introduces several key enhancements compared to its predecessor, v0.21.0 [Source event].
- Kimi CLI Integration: Users can now install and run the Kimi CLI directly through Ollama [Source event]. The Kimi CLI, with Kimi K2.6, excels at long-horizon agentic execution tasks using a multi-agent system [Source event].
- MLX Runner Logprobs Support: The MLX runner now supports logprobs for compatible models [Source event].
- Faster MLX Sampling: MLX sampling is faster due to fused top-P and top-K in a single sort pass [Source event]. Repeat penalties are also applied within the sampler [Source event].
- Improved MLX Prompt Tokenization: Prompt tokenization for MLX is enhanced by moving it into request handler goroutines [Source event].
- Better MLX Thread Safety: Array management in MLX now has improved thread safety [Source event].
- GLM4 MoE Lite Performance: GLM4 MoE Lite performance is improved with a fused sigmoid router head [Source event].
- Fixed Model Picker: The macOS app’s model picker no longer shows stale models after switching chats [Source event].
- Fixed Gemma 4 Structured Outputs: Structured outputs for Gemma 4 are fixed when
think=false[Source event].
How does Ollama v0.21.1 work
Ollama v0.21.1 integrates new features and optimizations into its existing framework [Source event].
- Kimi CLI Launch: Users launch the Kimi CLI via the
ollama launch kimicommand, specifying a model likekimi-k2.6:cloud[Source event]. - Multi-Agent Execution: The Kimi CLI, powered by Kimi K2.6, utilizes a multi-agent system for complex, long-horizon tasks [Source event].
- MLX Runner Enhancements: The MLX runner processes models with improved sampling efficiency [Source event]. Fused top-P and top-K operations streamline the sampling process [Source event].
- Logprobs Calculation: Compatible models benefit from logprobs support within the MLX runner [Source event].
- Optimized Tokenization: Prompt tokenization is handled by dedicated request handler goroutines, improving efficiency [Source event].
- Thread-Safe Array Management: MLX ensures better thread safety for array operations, contributing to stability [Source event].
- GLM4 MoE Lite Optimization: A fused sigmoid router head boosts the performance of GLM4 MoE Lite models [Source event].
Benchmarks and evidence
Specific benchmark results for Ollama v0.21.1 are not yet disclosed.
| Feature/Improvement | Claimed Benefit | Evidence Source |
|---|---|---|
| Kimi CLI with Kimi K2.6 | Excels at long horizon agentic execution tasks through a multi-agent system. | [Source event] |
| Faster MLX sampling | Achieved with fused top-P and top-K in a single sort pass. | [Source event] |
| Improved MLX prompt tokenization | By moving tokenization into request handler goroutines. | [Source event] |
| GLM4 MoE Lite performance | Improved with a fused sigmoid router head. | [Source event] |
Who should care
Builders
Builders should care about Ollama v0.21.1 for its enhanced MLX capabilities and the new Kimi CLI [Source event]. The improved MLX sampling and logprobs support offer more control and speed for model development [Source event]. The Kimi CLI provides a robust multi-agent system for complex automation tasks [Source event].
End users
End users of Ollama’s macOS app will benefit from the fixed model picker, ensuring a smoother experience [Source event]. Users interacting with Gemma 4 models will find structured outputs more reliable [Source event].
How to use Ollama v0.21.1 today
To use the Kimi CLI with Ollama v0.21.1, you can launch it via the command line [Source event].
Launch Kimi CLI:
ollama launch kimi --model kimi-k2.6:cloudThis command installs and runs the Kimi CLI, leveraging the kimi-k2.6:cloud model [Source event].
Risks, limits, and myths
- Myth: Ollama v0.21.1 is a standalone game. Fact: Ollama v0.21.1 is an AI platform update, not related to games like “Simple Days v0.21.1” by Mega Lono [Source event, 1, 2, 3, 4].
- Limit: Kimi CLI requires specific models. The Kimi CLI is demonstrated with
kimi-k2.6:cloudfor optimal agentic execution [Source event]. - Risk: Compatibility with older models. While MLX improvements are noted, specific compatibility details for all older models are not explicitly stated [Source event].
FAQ
- What is the release date of Ollama v0.21.1?
- Ollama v0.21.1 was released on [Source event].
- What is Kimi CLI in Ollama v0.21.1?
- Kimi CLI is a command-line interface that can be installed and run through Ollama, excelling at long-horizon agentic execution tasks [Source event].
- Which model does Kimi CLI use for agentic tasks?
- The Kimi CLI with Kimi K2.6 excels at agentic execution tasks through a multi-agent system [Source event].
- How does Ollama v0.21.1 improve MLX performance?
- Ollama v0.21.1 improves MLX performance with faster sampling, logprobs support, and better prompt tokenization [Source event].
- Are there any fixes for Gemma 4 in this release?
- Yes, structured outputs for Gemma 4 are fixed when
think=falsein Ollama v0.21.1 [Source event]. - What is the improvement for GLM4 MoE Lite?
- GLM4 MoE Lite sees performance improvement with a fused sigmoid router head in Ollama v0.21.1 [Source event].
- Does Ollama v0.21.1 fix issues in the macOS app?
- Yes, the model picker in the macOS app no longer shows stale models after switching chats [Source event].
- Is Ollama v0.21.1 related to any games?
- No, Ollama v0.21.1 is an AI platform update and is not related to games like “Simple Days v0.21.1” [1, 2, 3, 4].
Glossary
- Agentic Execution
- The capability of an AI system to perform complex, multi-step tasks autonomously, often involving planning and decision-making [Source event].
- MLX Runner
- A component within Ollama responsible for executing machine learning models, particularly those optimized for Apple silicon [Source event].
- Logprobs
- Logarithmic probabilities, often used in language models to represent the likelihood of generated tokens [Source event].
- Top-P Sampling
- A text generation technique that samples from the smallest set of tokens whose cumulative probability exceeds a threshold P [Source event].
- Top-K Sampling
- A text generation technique that samples from the K most likely next tokens [Source event].
- Goroutines
- Lightweight, concurrently executing functions in the Go programming language, used for efficient task handling [Source event].
- GLM4 MoE Lite
- A specific variant of the GLM4 model architecture, likely a Mixture-of-Experts (MoE) model, optimized for efficiency [Source event].
