Instructions to use myjellylabs/Krypton-Fly-2.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use myjellylabs/Krypton-Fly-2.0 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="myjellylabs/Krypton-Fly-2.0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("myjellylabs/Krypton-Fly-2.0")
model = AutoModelForCausalLM.from_pretrained("myjellylabs/Krypton-Fly-2.0", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use myjellylabs/Krypton-Fly-2.0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "myjellylabs/Krypton-Fly-2.0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "myjellylabs/Krypton-Fly-2.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/myjellylabs/Krypton-Fly-2.0

SGLang

How to use myjellylabs/Krypton-Fly-2.0 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "myjellylabs/Krypton-Fly-2.0" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "myjellylabs/Krypton-Fly-2.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "myjellylabs/Krypton-Fly-2.0" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "myjellylabs/Krypton-Fly-2.0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use myjellylabs/Krypton-Fly-2.0 with Docker Model Runner:
```
docker model run hf.co/myjellylabs/Krypton-Fly-2.0
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Krypton Fly 2.0 (`Krypton`)

Krypton Fly 2.0 is a 2.5B-parameter autonomous agentic, deep-reasoning (<think>...</think>), and software-engineering language model trained and released by Jelly Lab Game (myjellylabs).

Built on top of myjellylabs/Krypton-Fly-1.0 (LlamaForCausalLM, 42 layers, GQA, 131,072-token context window, 16,384 max output tokens), Krypton Fly 2.0 specializes in:

100% Native XML Tool-Calling (<function name="..."><param name="...">...</param></function>): Multi-turn repository debugging, MCP tool orchestration, and automatic error recovery with CDATA support for raw code and SQL parameters.
2026 Frontier Deep <think> Reasoning: Calibrated on verified <think> traces across Claude Opus 4.7 / 4.6, Fable 5.1, Gemini 3.1 Pro, Kimi K3 / K2.6, GPT-5.4 / 5.2, DeepSeek V4 / V3.2 Speciale, and Kimi-Cyber / CyberSec-Reasoning-Premium with strict fast-mode loss masking so enable_thinking=True and enable_thinking=False both work cleanly.
2026 Software Engineering & Agentic Debugging: Grounded repository repair, code synthesis, and algorithmic verification.

Training Mixture (4017 Curated Samples, ~8.00M Tokens)

Sector	Samples	Verified Source Dataset(s)
`Claude-Opus4.7-Fable5-CoT`	669	Met4physics/claude-opus-4-8-xhigh-reasoning-8.7k, 11-47/claude_opus_4.8_max_thinking_5k_v2, eddieran/opus-4.7-reasoning-cot, TeichAI/lordx64-claude-opus-4.7-max-cleaned, qsardor/Claude-Sonnet-Opus, TeichAI/Claude-Sonnet-4.6-Reasoning-1100x, saidutta69/fable-5.1-premium, saidutta69/fable-5-premium-v2
`Frontier-Coding-2026`	575	PawanKrd/claude-fable-5-code, beyoru/kimi-k3-distillation (sft_code), TeichAI/gpt-5-codex-1000x, TeichAI/deepseek-v3.2-speciale-OpenCodeReasoning-3k, nvidia/OpenCodeReasoning
`Tool-Use`	550	openbmb/UltraData-SFT-Agent-2609 (Tool-Use)
`Fable5.1-KimiK3-Sol-Agent-XML`	542	saidutta69/fable-5.1-premium-v2, greghavens/fable-5-coding-and-debugging-traces, greghavens/kimi-k3-coding-and-debugging-traces, beyoru/kimi-k3-distillation (sft), Crownelius/GPT-5.6-Sol-Luna-Terra-Traces, echel0nn1881/kimi-cyber-reasoning (XML tools)
`KimiK3-GPT61Sol-Astra-Cyber`	527	FlaneAI/GPT-6.1-Sol-Problem-solving-Reasoning-Distilled, saidutta69/CyberSec-Reasoning-Premium, echel0nn1881/kimi-cyber-reasoning, beyoru/kimi-k3-distillation (sft_code), uniquealexx/Kimi-K2.6-Thinking-200x, Roman1111111/gpt-5.4-step-by-step-reasoning
`Code-Agent`	420	openbmb/UltraData-SFT-Agent-2609 (Code-Agent)
`FastMode-LongOutput-CodeCyber`	185	PawanKrd/claude-fable-5-code, saidutta69/CyberSec-Reasoning-Premium, dpevzner/Cybersecurity_Reasoning_Dataset, Roman1111111/gpt-5.4-step-by-step-reasoning (Fast-Mode Extended Output)
`AgentForge-ErrorRecovery-XML`	170	voxozi/agentforge-multiturn-toolcall
`Hermes-SelfCorrection-XML`	170	DJLougen/hermes-agent-traces-filtered
`Frontier-CyberSec-VulnAudit-2026`	139	saidutta69/CyberSec-Reasoning-Premium, echel0nn1881/kimi-cyber-reasoning, p-research/kimi-k3-cyber-reasoning-distill, davidquicast/vulnerability-intelligence-diagrammatic-reasoning, dpevzner/Cybersecurity_Reasoning_Dataset, me-aas/pentesting-explanations
`identity_anchor`	70	Krypton Fly 2.0 Identity Anchor (Jelly Lab Game)

Evaluation Summary (`myjellylabs/Krypton-Fly-1.0` vs Krypton Fly 2.0)

Metric	`myjellylabs/Krypton-Fly-1.0`	Krypton Fly 2.0
Identity (Krypton / Krypton Fly 2.0)	0.60	0.70
Native XML Tool-Call Validity (`<function><param>`)	1.00	1.00
Native XML Tool-Name Exact Match	0.54	0.54
Math & Coding Accuracy (Fast Mode)	1.00	1.00
Deep `<think>` Mode Closes `</think>` Properly	0.90	1.00
Deep `<think>` Mode Final Answer Accuracy	0.90	1.00
Chat Completion Within Token Budget	1.00	0.93
Avg Fast-Mode Output Length (chars)	1442.33	2138.67
Avg Post-`</think>` Output Length (chars)	1004.90	465.60
Core Sanity Checks	1.00	1.00
Format Constraint Following	1.00	1.00
Version 2.0 Self-Awareness	0.00	0.00

Quickstart Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "myjellylabs/Krypton-Fly-2.0"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "Write a thread-safe LRU cache in Python with O(1) get and put."}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.7, top_p=0.95)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

License

Released under Apache-2.0 by Jelly Lab Game (myjellylabs). See LICENSE and NOTICE.

Downloads last month: 19

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for myjellylabs/Krypton-Fly-2.0

Quantizations

1 model