Jugurtha-4B (GGUF)

Quantized GGUF weights for Jugurtha-4B, the first instruction-tuned foundation model for Kabyle (Taqbaylit, kab, Latin script) with native chain-of-thought reasoning.

Optimized for consumer hardware, edge devices, LM Studio, and Ollama.

Provided Quantizations

---

Direct Usage with Ollama

Run directly from Hugging Face without manual downloads:

# Direct fast responses (Thinking OFF)
ollama run hf.co/agbalu/Jugurtha-4B-GGUF --think=false

# Full Chain-of-Thought reasoning in Taqbaylit (Thinking ON)
ollama run hf.co/agbalu/Jugurtha-4B-GGUF --think=true

Inside an interactive Ollama session, you can toggle thinking at any time:

• /set nothink — immediate, concise answers
• /set think — detailed internal reasoning in Taqbaylit

---

Direct Usage with LM Studio

1. Open LM Studio.
2. Search for agbalu/Jugurtha-4B-GGUF.
3. Choose your quantization (Q4_K_M recommended) and click Download.
4. Click Load Model. Ensure GPU Offload is set to Max for peak speed (40–70 tok/s on Apple Silicon / CUDA).
5. Toggle the Thinking / Reasoning switch in the sidebar to turn reasoning on or off.

---

Direct Usage with llama.cpp

For fast multi-threaded GPU inference:

# Thinking OFF (direct answer)
llama-cli -m Jugurtha-4B-Q4_K_M.gguf -ngl 99 -t 8 \
    -p "<|im_start|>user\nSegzi-yi-d s wegzul d acu i d Agbalu.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n" \
    -n 512

# Thinking ON (chain-of-thought in Taqbaylit)
llama-cli -m Jugurtha-4B-Q4_K_M.gguf -ngl 99 -t 8 \
    -p "<|im_start|>user\nAcḥal i d 25 x 4?<|im_end|>\n<|im_start|>assistant\n<think>\n" \
    -n 512

---

Main Model & Details

For full model benchmarks, training recipe, and architecture details, refer to the main repository: agbalu/Jugurtha-4B.

Licence

Apache-2.0