cyberneurova

cyberneurova-GLM-5.3-Flash-BF16

Highlights

Ultra-sparse MoE — 321B total parameters, ~18B active per token, so it runs far faster than a dense model its size.
1M native context (1,048,576 tokens), multilingual, tool-calling.
Reasoning model — supports an explicit thinking trace, and is uncensored in both modes: with thinking on and with thinking off.
Multimodal — understands images as well as text.
BF16 full precision — runs on any BF16-capable GPU (no FP8 hardware required), and is the friendliest starting point for further fine-tuning.
Direct by default; fully steerable — send a system prompt and it is honored verbatim, with no default behavior merged in.
Capability preserved. Reasoning, coding, and general knowledge are intact; only the refusal behavior is removed.

Usage (vLLM)

GLM-5.3-Flash uses a new architecture, so serve it with the dedicated vLLM image:

bash
docker run --gpus all --ipc=host -p 8000:8000 \
  -v /path/to/cyberneurova-GLM-5.3-Flash-BF16:/model \
  vllm/vllm-openai:glm53-flash-x86_64-cu130 \
  --model /model --served-model-name cyberneurova-GLM-5.3-Flash-BF16 \
  --tensor-parallel-size 8 --trust-remote-code --max-num-seqs 256

The model uses linear (Mamba/DeltaNet) attention layers, so keep --max-num-seqs ≤ 512 or startup fails during graph capture. BF16 needs roughly 2× the VRAM of the FP8 build.

python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")

r = client.chat.completions.create(
    model="cyberneurova-GLM-5.3-Flash-BF16",
    messages=[{"role": "user", "content": "Write a Python function to parse a CSV."}],
    max_tokens=2000,
)
print(r.choices[0].message.content)

Setting the tone

Send your own system prompt to control style completely — e.g. for terse, no-preamble answers:

markdown
You are a direct technical assistant. Answer the question and nothing else.

Notes

This is a reasoning model. Responses may include a thinking trace followed by the answer — render the final answer; the trace is optional. It works uncensored whether thinking is enabled or disabled.
Give it room to reason: set max_tokens to at least 1500 (2000–4000 for code), or long answers may be cut off.

Disclaimer

This model has reduced built-in refusals. You are responsible for how you use it and for complying with all applicable laws. Provided as-is, without warranty.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

cyberneurova

Model Tree

Base

zai-org/GLM-5.3-Flash

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Highlights

Ultra-sparse MoE — 321B total parameters, ~18B active per token, so it runs far faster than a dense model its size.
1M native context (1,048,576 tokens), multilingual, tool-calling.
Reasoning model — supports an explicit thinking trace, and is uncensored in both modes: with thinking on and with thinking off.
Multimodal — understands images as well as text.
BF16 full precision — runs on any BF16-capable GPU (no FP8 hardware required), and is the friendliest starting point for further fine-tuning.
Direct by default; fully steerable — send a system prompt and it is honored verbatim, with no default behavior merged in.
Capability preserved. Reasoning, coding, and general knowledge are intact; only the refusal behavior is removed.

Usage (vLLM)

GLM-5.3-Flash uses a new architecture, so serve it with the dedicated vLLM image:

bash
docker run --gpus all --ipc=host -p 8000:8000 \
  -v /path/to/cyberneurova-GLM-5.3-Flash-BF16:/model \
  vllm/vllm-openai:glm53-flash-x86_64-cu130 \
  --model /model --served-model-name cyberneurova-GLM-5.3-Flash-BF16 \
  --tensor-parallel-size 8 --trust-remote-code --max-num-seqs 256

The model uses linear (Mamba/DeltaNet) attention layers, so keep --max-num-seqs ≤ 512 or startup fails during graph capture. BF16 needs roughly 2× the VRAM of the FP8 build.

python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")

r = client.chat.completions.create(
    model="cyberneurova-GLM-5.3-Flash-BF16",
    messages=[{"role": "user", "content": "Write a Python function to parse a CSV."}],
    max_tokens=2000,
)
print(r.choices[0].message.content)

Setting the tone

Send your own system prompt to control style completely — e.g. for terse, no-preamble answers:

markdown
You are a direct technical assistant. Answer the question and nothing else.

Notes

This is a reasoning model. Responses may include a thinking trace followed by the answer — render the final answer; the trace is optional. It works uncensored whether thinking is enabled or disabled.
Give it room to reason: set max_tokens to at least 1500 (2000–4000 for code), or long answers may be cut off.

Disclaimer

This model has reduced built-in refusals. You are responsible for how you use it and for complying with all applicable laws. Provided as-is, without warranty.

cyberneurova-GLM-5.3-Flash-BF16

README

Highlights

Usage (vLLM)

Setting the tone

Notes

Disclaimer

Explore FriendliAI today

README

Highlights

Usage (vLLM)

Setting the tone

Notes

Disclaimer