Naim 9B · v2
Naim is an autonomous AI coding agent created by ABDESSEMED Mohamed (redhamohamed).
Lineage: Mimo → Naim
Mimo (redhamohamed/mimo-v5-gguf · 8B · by ABDESSEMED Mohamed)
│ identity, name, spirit: a private, local assistant that belongs to its creator
▼
Naim (redhamohamed/naim · Naim architecture · 9B · autonomous coding agent)
Naim is the successor of Mimo, Mohamed's earlier model (not related to Xiaomi's MiMo). Naim carries Mimo's identity and purpose forward as an autonomous coding agent.
Naim architecture
| Parameters | ~9B |
| Context | up to 262,144 tokens |
| Modes | reasoning (<think>) and direct answer, tool calling, vision (with naim-mmproj-f16.gguf) |
| Languages | French, English and many others |
model_type | naim (NaimForConditionalGeneration) |
How Naim works
Naim follows an agent loop on every task:
- Understand: restate the goal, read the relevant code, list constraints.
- Plan: split the work into small verifiable steps.
- Implement: clean, idiomatic code that matches the existing style.
- Verify: run tests or scripts, read errors, fix.
- Report: what changed, what was verified, what's left.
Files
| File | Use |
|---|---|
*.safetensors + *_naim.py | Transformers (bf16, trust_remote_code=True) |
naim-Q4_K_M.gguf | LM Studio / Ollama / llama.cpp: recommended (~5.6 GB) |
naim-Q8_0.gguf | Higher quality (~9.5 GB) |
naim-mmproj-f16.gguf | Vision projector for LM Studio / llama.cpp (lets Naim see images) |
Usage
Ollama
ollama run hf.co/redhamohamed/naim:Q4_K_M
LM Studio: search redhamohamed/naim and download naim-Q4_K_M.gguf (+ naim-mmproj-f16.gguf for images).
llama.cpp (with vision):
llama-server -m naim-Q4_K_M.gguf --mmproj naim-mmproj-f16.gguf
Transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("redhamohamed/naim", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("redhamohamed/naim", trust_remote_code=True, dtype=torch.bfloat16)
msgs = [{"role": "user", "content": "Qui es-tu ?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
out = model.generate(**ids, max_new_tokens=512)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))
Pass enable_thinking=False to apply_chat_template for direct answers without the reasoning phase.
Recommended sampling for code: temperature 0.3, top_p 0.95, top_k 20, presence_penalty 0 (a presence penalty breaks code: braces and names must repeat).
What's new in v2
- No more canned self-introductions: Naim answers "merci" like a normal assistant and only presents itself when asked.
- Answers distilled to keep the full coding skill (complete, compilable code with every import).
- Knows its identity (creator, Mimo lineage, Naim architecture) without inventing details.
License
Apache 2.0. Naim is a fine-tune of an open-weight Apache-2.0 base model; see the LICENSE file.