Yusiko/khazri-2

Khazri 2 Preview

Khazri 2 Preview is the flagship open-weight release in the current Khazri family. It combines the Khazri model with the project's SYNAPSE architecture and route-aware inference system for research, experimentation and controlled integration.

Hugging Face · Khazri · Contact

Preview release: Validate this model on your own tasks before production use. It should not be the sole basis for high-impact decisions.

At a glance

Item	Detail
Model	Khazri 2 Preview
Family position	Flagship model in the Khazri 2 generation
Parameters	~250M; the release config is the final source of truth
Status	Open weights on Hugging Face
Architecture	Khazri model with SYNAPSE components
Safe export context limit	512 tokens
Default maximum generation	128 new tokens
Loading	Transformers with custom model code
Web-search route	Disabled by default

SYNAPSE-aware inference

The release includes the configuration, tokenizer, safetensor weights and the custom SYNAPSE model files required for the inference stack:

config.json
generation_config.json
model.safetensors
tokenizer.json
tokenizer_config.json
modeling_synapse.py
synapse_controller.py
inference.py

The route system supports:

• NSR_CALCULATOR for safe numeric calculation;
• MATH_SOLVER for symbolic mathematics;
• TMS_CONTEXT for context-grounded responses;
• UQM_ABSTAIN when information is insufficient or inappropriate;
• WEB_SEARCH, disabled by default; and
• MODEL_FALLBACK for standard model generation.

Routing is implementation-dependent. Inspect the custom Python files before enabling remote code and pin a specific model revision in production.

Training-data documentation

Khazri 2 Preview does not state a source-level corpus mix in this model card. Do not assume that it uses the same sources, proportions or post-training process as Khazri 2 Mini.

For a particular release, use the data documentation and model revision published with the Khazri 2 Hugging Face repository. A complete data card should identify source datasets and licences, language coverage, filtering and deduplication steps, pre-training and instruction-tuning material, known limitations and revision-specific evaluation.

Installation

pip install -U torch transformers accelerate safetensors

Quick start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Yusiko/khazri-2"

# Inspect the model repository before enabling custom remote code.
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    use_fast=True,
)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

prompt = "Summarize the following text in two concise sentences: ..."
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=384)
inputs = {key: value.to(model.device) for key, value in inputs.items()}

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

For this preview export, reserve generation tokens within the 512-token total context limit. For example, use up to 384 input tokens with 128 new tokens.

Compact-model comparison

The following values are project-provided internal results for Khazri 2 Preview across selected compact-model tests. Higher is better.

Model	Parameters	Context extraction	Mixed speed/proxy	Arithmetic	Word problems	Abstention
Khazri 2 Preview	~250M	100%	62%	99%	99%	97.4%
Gemma 3	270M	100%	36%	0%	0%	18%
Qwen 2.5	0.5B	89%	45%	14%	28%	46%
Pythia	160M	22%	10%	0%	2%	1%

These figures are not independently audited. They describe the reported test configuration only and do not establish general superiority or predict performance on every downstream task. Publish prompts, task definitions, scoring rules, model versions, seeds, hardware and full evaluation assets with future public benchmark claims.

Responsible use

Khazri 2 Preview may hallucinate, make reasoning errors, reflect training-data bias or abstain incorrectly. Use retrieval, verification and human review for consequential work. Do not rely on outputs alone for medical, legal, financial, security, safety, education, employment or other high-impact decisions.

Release notes

Use the files, revision and licence stated on the Hugging Face model page as the source of truth for a specific release.

Roadmap

Khazri 2 Preview represents the current flagship stage. The next goal is Khazri 3: a model with a larger parameter count and stronger results.

Contact

To discuss research, integration or partnerships, visit khazri.dev or email contact@khazri.dev.