Apeiron 1.1 Preview

Apeiron 1.1 Preview by Forever AI (formerly Klox AI), Ape's line. An early, focused release designed to establish identity, language capabilities, and tool-use mechanics, prioritizing behavioral alignment over raw factual knowledge. Built entirely from scratch by our research team.

Everything you need is in one place:

• Full Model Weights (model.safetensors) — for Python integration;
• Ready-to-use GGUF files for llama.cpp / Ollama / LM Studio;
• Modelfile for Ollama;
• Training Dataset: forever-ai/apeiron-1.1-preview-set.

Benchmarks

Core Capabilities

• Identity Anchoring: Firmly knows it is Apeiron 1.1, remembers its predecessor Apeiron 1.0, and identifies with the Ape's family by Forever AI.
• Language Adaptation: Naturally responds in the user's language (tested on Russian, English, and others). The early-stage artifact of appending "Reply in Russian" has been completely fixed.
• Professional Tone: Maintains a concise, business-oriented tone.
• Native Tool Calling: Executes tool calls in our proprietary format (<tool_call> / <tool_response>).
  • Live-tested: Web search (Bing), page fetching (real HTTP 200), calculator, file reading, and time.
  • Multi-step reasoning: Capable of chaining actions (e.g., search → read page → answer).

Launch without Python (llama.cpp / Ollama / LM Studio)

llama.cpp:

llama-cli -m apeiron-1.1-preview-q4_k_m.gguf -p "Who are you?" -n 128

Ollama:

ollama create apeiron -f Modelfile
ollama run apeiron

LM Studio: Simply open the file as a model — the chat template is already embedded inside.

Launch in Python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "forever-ai/apeiron-1.1-preview"

tok = AutoTokenizer.from_pretrained(MODEL_ID)
# Loading our full proprietary weights directly
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

SYS = ("You are Apeiron 1.1 by Forever AI (formerly Klox AI), an agent of the Ape's family. "
       "Before you, Apeiron 1.0 was released—the first model in the line. "
       "Your response language must match the user's language.")
msgs = [{"role": "system", "content": SYS}, {"role": "user", "content": "Who are you?"}]
inp = tok.apply_chat_template(msgs, add_generation_prompt=True,
                              return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=160, do_sample=False)
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))

Scope of this Preview

This is a functional preliminary release. While it excels at identity, language matching, and tool execution, users should note:

• Knowledge: Factual knowledge is intentionally limited in this iteration (deep "smartness" will arrive in the next major training phase).
• Tooling in GGUF: Tools (search, calculator, etc.) do not natively execute within the GGUF files. They require our external Python wrapper (apeiron_agent.py).
• Testing: Long-context performance and safety alignment are still being evaluated and have not been fully tested yet.