Model releases · 11 August 2026
Two small models trained specifically for French, released under Apache 2.0 with quantised builds small enough to run on a phone — and an unusual detail in what they were built from.
| Released by | Kurakura AI (kurakurai) |
|---|---|
| Published | August 2026 |
| Variants | Luth-2-0.8B and Luth-2-2B |
| Parameters | 1.88B text-only for the 2B variant |
| Base model | Qwen3.5-2B |
| Architecture | Qwen3_5ForConditionalGeneration (qwen3_5) |
| Layers / hidden size | 24 / 2048 (2B variant) |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 tokens |
| Language | French |
| Licence | Apache 2.0 |
| Quantised builds | GGUF published by the lab: Q4_K_M 1.27 GB, Q8_0 2.01 GB, F16 3.78 GB |
| Hosted at | kurakurai/Luth-2-2B and -0.8B on Hugging Face |
Luth-2 is a pair of small language models trained for French. The lab claims state of the art in French for their size class across mathematics, code, instruction following, general knowledge and tool calling, measured on their own selected French benchmarks, and positions the models explicitly for local and on-device deployment.
Both variants are non-reasoning models — no extended chain-of-thought mode — which is a deliberate choice at this size.
The detail most coverage will skip: Luth-2-2B's configuration declares Qwen3_5ForConditionalGeneration, with Qwen3.5-2B listed as the base model. That places this release on the Qwen 3.5 generation rather than an older base.
It carries a consequence the lab flags directly in the model card: Luth-2-2B inherits the vision-language architecture of Qwen3.5-2B, but was not trained on vision data. The lab explicitly recommends against using it for vision tasks. The 1.88B parameter figure is text-only.
The recipe is a 3-billion-token French supervised fine-tuning mixture, followed by multi-domain on-policy distillation. Both post-training datasets have been published alongside the models, as has the training code.
That distillation approach is the same family of technique named in the Motif 3 release published the same week — on-policy distillation from teacher models appearing at both ends of the size range, from a 314B MoE down to a sub-2B French specialist.
Unusually for a small lab, the release is complete rather than weights-only: both model sizes, GGUF conversions published by the lab itself rather than left to the community, the SFT and RL datasets, the training code on GitHub, and a French LLM leaderboard space for the evaluation.
The GGUF builds start at 1.20 GB for Q4_0 and 1.27 GB for Q4_K_M on the 2B variant, with the 0.8B variant smaller again.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | GGUF builds are published by the lab across Q4_0, Q4_K_M, Q5_K_M, Q6_K, Q8_0 and F16. |
| transformers | Yes | Standard configuration; requires a version with Qwen3.5 support. |
| MLX | No | No MLX conversion published by the lab at release. |
| Vision input | No | The architecture inherits vision support from Qwen3.5-2B, but the model was not trained on vision data and the lab recommends against it. |
A pair of small language models trained specifically for French, at 0.8B and 2B parameters, released by Kurakura AI under Apache 2.0 with quantised GGUF builds published alongside the weights.
Qwen3.5-2B. The 2B variant declares the Qwen3.5 architecture and lists Qwen3.5-2B as its base model.
No. It inherits the vision-language architecture of its Qwen3.5 base but was not trained on vision data, and the lab explicitly recommends against using it for vision tasks.
The 2B variant is 1.27 GB at Q4_K_M and 1.20 GB at Q4_0, published by the lab. The 0.8B variant is smaller still.
Yes. Both the supervised fine-tuning and reinforcement learning post-training datasets are published on Hugging Face, along with the training code on GitHub.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026