Instructions to use tablinum/Moon-Beam-8B-A1B-Preview with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tablinum/Moon-Beam-8B-A1B-Preview with Transformers:

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

pipe = pipeline("text-generation", model="tablinum/Moon-Beam-8B-A1B-Preview", trust_remote_code=True)
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("tablinum/Moon-Beam-8B-A1B-Preview", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("tablinum/Moon-Beam-8B-A1B-Preview", trust_remote_code=True, 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 tablinum/Moon-Beam-8B-A1B-Preview with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tablinum/Moon-Beam-8B-A1B-Preview"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tablinum/Moon-Beam-8B-A1B-Preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tablinum/Moon-Beam-8B-A1B-Preview

SGLang

How to use tablinum/Moon-Beam-8B-A1B-Preview 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 "tablinum/Moon-Beam-8B-A1B-Preview" \
    --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": "tablinum/Moon-Beam-8B-A1B-Preview",
		"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 "tablinum/Moon-Beam-8B-A1B-Preview" \
        --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": "tablinum/Moon-Beam-8B-A1B-Preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tablinum/Moon-Beam-8B-A1B-Preview with Docker Model Runner:
```
docker model run hf.co/tablinum/Moon-Beam-8B-A1B-Preview
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Moon-beam 8B A1B · BF16 · preview

Moon-beam is a mixture-of-experts model with 8B total parameters, about 1B active per token, and a 131,072-token context window. It pairs inclusionAI’s Ling-3.0-tiny with Tablinum’s searchable knowledge library, the engram. Round 1b was fine-tuned with QLoRA on retrieved passages, math problems, tool calls, and multi-turn conversations. Model and evaluation notes

On the preliminary November 4 evaluation, Moon-beam scored 96.2% on 400 GSM8K questions with thinking on. On EngramBench it answered 440 of 503 questions correctly (87.5%) with thinking off and passages from each question’s own table.

Preview. These weights are the release candidate (round 1b). They run today with the Tablinum-2 library. Tablinum-3, about twice the library, is still being built and tested; when it is, these cards and the engram pack will be updated and the -Preview suffix dropped. Numbers marked preliminary are being re-measured with the model-card protocol (temperature 1.0, top_p 0.95, top_k 20, up to 32K new tokens, full 131K window).

Run it

With the library (recommended): the GGUF repo runs with one command, terrabeam run tablinum/Moon-Beam-8B-A1B-GGUF-Preview. Or serve these weights and put terrabeam in front:

vllm serve tablinum/Moon-Beam-8B-A1B-Preview --trust-remote-code --max-model-len 131072      # or SGLang

Add the library. Moon-beam is trained to answer from retrieved passages. Put terrabeam in front of your server and every request gets them, in the exact layout the model was trained on:

terrabeam proxy --upstream http://127.0.0.1:8000 --format tablinum --memory long-term   # then use http://127.0.0.1:8090/v1

Transformers (4.57; the model code ships in this repo):

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("tablinum/Moon-Beam-8B-A1B-Preview", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("tablinum/Moon-Beam-8B-A1B-Preview", trust_remote_code=True, torch_dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "What is 17 * 23?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=2048, do_sample=True, temperature=1.0, top_p=0.95, top_k=20)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Files

File	What it is
`model-0000*-of-00004.safetensors`	BF16 weights (16 GB), Ling-3.0-tiny architecture (`bailing_hybrid`: KDA + MLA attention, MoE)
`modeling_bailing_moe_v3.py`, `configuration_bailing_moe_v3.py`	Model code (`trust_remote_code=True`)
`mtp/mtp_layer.safetensors`	The trained MTP draft layer (`model.layers.24.*`, 0.63 GB) for speculative decoding
`chat_template.jinja`, `generation_config.json`	Thinking on by default; temperature 1.0, top_p 0.95, top_k 20

Other formats: GGUF · FP8 · MLX · MLX 8-bit

How much does retrieval help?

With thinking on, retrieved passages raised SimpleQA from 28 to 593 correct answers out of 1,000. The same comparison on EngramBench went from 65 to 431 correct out of 503. Claude Haiku graded the answers, with partial answers and abstentions counted separately from correct answers.

Read the figures as text

Test (questions)	Thinking on · closed → retrieved	Thinking off · closed → retrieved
EngramBench (503)	12.9% → 85.7%	10.1% → 87.5%
SimpleQA (1000)	2.8% → 59.3%	2.6% → 57.5%
AA-Omniscience (200)	10.0% → 26.5%	5.0% → 26.5%

The untrained Ling base also reached 59.2% on SimpleQA with the same passages and thinking on, compared with 59.3% for round 1b. The base used IQ3 and earlier settings. The clearest gain in these runs comes from giving the model relevant passages. Graded counts and settings

Reasoning and follow-ups

Across 150 question pairs, Moon-beam answered 94.0% of first questions and 93.3% of second questions correctly in the same conversation with thinking on.

Read the figures as text

Test · sample size	Thinking on	Thinking off
GSM8K · 400	96.2%	92.5%
Multi-turn, first · 150	94.0%	86.7%
Multi-turn, second · 150	93.3%	88.7%
MMLU-Pro · 700	63.9%	44.0%

The thinking-off column used greedy decoding. On MMLU-Pro, 70 of 700 thinking-on answers reached the 30,000-token output cap. The sampled base also scored 96.2% on GSM8K. The release evaluation uses revised sampling and a 32K output allowance. Preliminary results and protocol

Talking a customer through a fix

On tau2-bench — multi-turn support conversations where Moon-beam is the agent, a Claude Haiku customer calls with a problem, and database and environment checks decide success — Moon-beam earned an average reward of 0.71 on 50 airline tasks and 0.80 on 114 retail tasks, with thinking on. The earlier as-shipped harness capped each turn and dropped the model's reasoning between steps, scoring 0.05 and 0.22 on the same domains; carrying the reasoning forward, in the layout Moon-beam was trained on, is what changed. Telecom and banking are being measured the same way.

tau2 domain · tasks	Corrected protocol	As-shipped harness
Airline · 50	0.711	0.045
Retail · 114	0.797	0.217
Telecom · 114	in progress	0.000
Banking	in progress	—

How small can the model file be?

Q4_K_M fits in a 4.94 GB file, 68.7% smaller than the BF16 GGUF. On 8,393 mixed chat, code, and math positions, it selected the same top token as BF16 97.65% of the time. On WikiText, agreement was 92.74%.

Read the figures as text

Build	File size	Mixed top-1 match	WikiText top-1 match
Q4_K_M	4.94 GB	97.65%	92.74%
Q6_K	7.23 GB	98.92%	96.74%
Q8_0	8.41 GB	99.18%	97.20%
MLX beam	4.92 GB	96.75%	90.92%
MLX 8-bit	8.87 GB	99.20%	97.40%

MLX beam uses 4-bit experts and 8-bit weights elsewhere. The MLX sizes above use the posted shard bytes in decimal GB. They correspond to 4.58 and 8.26 GiB, respectively. The full ladder includes all eleven GGUF and MLX measurements, with KL divergence and perplexity for both token sets. Quantization measurements · GGUF file sizes

Decode speed on a 12 GB GPU

On an RTX A2000 12 GB, the Q4_K_M model averaged 154.6 output tokens/s with the MTP draft head, up from 136.9 tokens/s alone. That is a 12.9% higher mean across three prompts with thinking off and greedy decoding. Two prompts sped up. The second slowed from 136.9 to 131.7 tokens/s.

Read the figures as text

Prompt	Trunk alone (tokens/s)	MTP n=2 (tokens/s)
1	137.3	176.5
2	136.9	131.7
3	136.6	155.7

The draft head is a separate 0.55 GB file. These measurements use a maximum of two draft tokens per step. Raw timings

Thanksgiving artwork

Moon-pie Rescue follows the Moon-beam cat as it catches a drifting pumpkin pie and offers you a slice. Open the 12-second MP4 · View the still banner

Small paws. Big thanks.

Model family

BF16 · GGUF · FP8 · MLX mixed-bit · MLX 8-bit

Credit and license

MIT, like the Ling-3.0-tiny base. Keep the notice; if Moon-beam or the Tablinum research helps you, a mention or a thank-you is appreciated. Library tables keep their own licenses: Tablinum-2. Runner: terrabeam (MIT).