Artificial Intelligence
Preferred on Google
Google News

llamafile v0.10.5 Ships With Big Local Models

Llamafile v0.10.5 adds support for two large local AI models, Ternary Bonsai 27B and Laguna-S-2.1, by updating its core llama.cpp integration.

8 min read
Screenshot of the llamafile v0.10.5 release notes on a computer screen.
Mozilla Blog

Visual TL;DR. llamafile v0.10.5 released includes Updated llama.cpp core. Outdated llama.cpp solves Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models like Ternary Bonsai 27B. Support for big models and Laguna-S-2.1 MoE. Support for big models leading to Efficient local AI. llamafile v0.10.5 released also features Packaging, docs upgrades.

  1. llamafile v0.10.5 released: new version of llamafile ships with updated core llama.cpp integration
  2. Outdated llama.cpp: older llamafile builds lacked support for new model architectures and quantization methods
  3. Updated llama.cpp core: integrates recent llama.cpp changes, enabling support for new model types
  4. Support for big models: now runs Ternary Bonsai 27B and Laguna-S-2.1 directly on local hardware
  5. Ternary Bonsai 27B: 6GB compressed Qwen3.6-27B model using ternary weights for efficiency
  6. Laguna-S-2.1 MoE: 118B parameter coding MoE model now accessible on consumer machines
  7. Efficient local AI: previously unwieldy models now fit and run efficiently on consumer-grade machines
  8. Packaging, docs upgrades: full details on announcement, including improved packaging and documentation
Visual TL;DR
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models leading to Efficient local AI includes enables leading to llamafile v0.10.5 released Updated llama.cpp core Support for big models Efficient local AI From startuphub.ai · The publishers behind this format
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models leading to Efficient local AI includes enables leading to llamafile v0.10.5released Updated llama.cppcore Support for bigmodels Efficient localAI From startuphub.ai · The publishers behind this format
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models leading to Efficient local AI includes enables leading to llamafile v0.10.5 released new version of llamafile ships withupdated core llama.cpp integration Updated llama.cpp core integrates recent llama.cpp changes,enabling support for new model types Support for big models now runs Ternary Bonsai 27B andLaguna-S-2.1 directly on local hardware Efficient local AI previously unwieldy models now fit and runefficiently on consumer-grade machines From startuphub.ai · The publishers behind this format
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models leading to Efficient local AI includes enables leading to llamafile v0.10.5released new version ofllamafile shipswith updated core… Updated llama.cppcore integrates recentllama.cpp changes,enabling support… Support for bigmodels now runs TernaryBonsai 27B andLaguna-S-2.1… Efficient localAI previously unwieldymodels now fit andrun efficiently on… From startuphub.ai · The publishers behind this format
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Outdated llama.cpp solves Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models like Ternary Bonsai 27B. Support for big models and Laguna-S-2.1 MoE. Support for big models leading to Efficient local AI. llamafile v0.10.5 released also features Packaging, docs upgrades includes solves enables like and leading to also features llamafile v0.10.5 released new version of llamafile ships withupdated core llama.cpp integration Outdated llama.cpp older llamafile builds lacked support fornew model architectures and quantizationmethods Updated llama.cpp core integrates recent llama.cpp changes,enabling support for new model types Support for big models now runs Ternary Bonsai 27B andLaguna-S-2.1 directly on local hardware Ternary Bonsai 27B 6GB compressed Qwen3.6-27B model usingternary weights for efficiency Laguna-S-2.1 MoE 118B parameter coding MoE model nowaccessible on consumer machines Efficient local AI previously unwieldy models now fit and runefficiently on consumer-grade machines Packaging, docs upgrades full details on announcement, includingimproved packaging and documentation From startuphub.ai · The publishers behind this format
Visual TL;DR, startuphub.ai llamafile v0.10.5 released includes Updated llama.cpp core. Outdated llama.cpp solves Updated llama.cpp core. Updated llama.cpp core enables Support for big models. Support for big models like Ternary Bonsai 27B. Support for big models and Laguna-S-2.1 MoE. Support for big models leading to Efficient local AI. llamafile v0.10.5 released also features Packaging, docs upgrades includes solves enables like and leading to also features llamafile v0.10.5released new version ofllamafile shipswith updated core… Outdatedllama.cpp older llamafilebuilds lackedsupport for new… Updated llama.cppcore integrates recentllama.cpp changes,enabling support… Support for bigmodels now runs TernaryBonsai 27B andLaguna-S-2.1… Ternary Bonsai27B 6GB compressedQwen3.6-27B modelusing ternary… Laguna-S-2.1 MoE 118B parametercoding MoE modelnow accessible on… Efficient localAI previously unwieldymodels now fit andrun efficiently on… Packaging, docsupgrades full details onannouncement,including improved… From startuphub.ai · The publishers behind this format

The latest release of llamafile, version 0.10.5, is now available, bringing enhanced support for running substantial AI models directly on personal hardware. This update tracks a more recent llama.cpp core, unlocking the ability to deploy models like the 6GB Ternary Bonsai 27B and the 118B parameter Laguna-S-2.1 coding MoE. You can find the full details on this announcement.

These models, previously available as GGUF weights, were inaccessible in older llamafile builds due to outdated llama.cpp support. The new version addresses this by incorporating architectures and quantization methods that allow these previously unwieldy models to fit and run efficiently on consumer-grade machines.

Shrinking Giants: Ternary Bonsai and Laguna-S-2.1

The inclusion of Ternary Bonsai 27B is particularly noteworthy. This model is a compressed version of Qwen3.6-27B from PrismML, employing ternary weights. This means each weight is restricted to {-1, 0, +1}, drastically reducing the data needed per weight to approximately 1.58 bits, down from the typical 16 bits. The result is a 27 billion parameter model that occupies only around 6GB of disk space. It's designed to maintain much of its original quality while being suitable for laptop use, and it supports multimodal capabilities, loading its vision component only when needed.

Laguna-S-2.1, developed by Poolside, is an open-weight coding model. While listed at 118 billion parameters, it functions as a Mixture-of-Experts (MoE). This architecture means only a fraction of its parameters, around 8B, are active for processing each token. This efficiency allows a quantized version to fit within consumer memory and operate at practical speeds. It's built for complex coding tasks and extended interactions, boasting a large context window. Although the BF16 weights support a 1 million token context, the quantized checkpoints are configured for 256K tokens.

StartupHub.ai data shows Poolside with a score of 77/100, placing it alongside competitors like Cursor (Anysphere) and Cognition AI. Poolside has also secured $1 billion in Series A funding in 2025, a verified financial milestone according to StartupHub.ai data.

The Sync Behind the Models

The ability to run these advanced models locally hinges on llamafile's constant synchronization with the rapidly evolving llama.cpp project. The release notes highlight three syncs in just two weeks, pushing the embedded llama.cpp through multiple upstream builds. Improvements to an agent skill that automates these syncs have accelerated the process, reducing the typical months-long delay between a model's release and its availability through llamafile. This focus on efficient packaging and integration is central to llamafile's value proposition.

Packaging and Documentation Upgrades

Beyond model support, llamafile v0.10.5 also introduces useful enhancements. The self-contained speech-to-text executable, transcribefile, now ships as a pre-built release artifact, removing the need for manual compilation. Documentation has also been improved, clarifying the help system, release binaries, and GPU support, including the Vulkan backend, thanks to community contributions.

This release signifies a continued push toward making powerful AI models accessible and usable on personal devices. The ability to run models of this scale locally, interactively, was a distant prospect just a year ago. Llamafile's ongoing development is key for developers and users seeking powerful AI without the reliance on cloud infrastructure.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.
#llamafile#llama.cpp#Ternary Bonsai 27B#Qwen3.6-27B#PrismML#Laguna-S-2.1#Poolside#Mixture-of-Experts#GGUF#transcribefile