Moon-beam 8B A1B · GGUF · 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 model-card protocol evaluation (temperature 1.0, top_p 0.95, top_k 20, up to 32K new tokens, the full 131K window), Moon-beam scored 95.5% on 400 GSM8K questions with thinking on. On EngramBench it answered 436 of 503 questions correctly (86.5%) with thinking on and passages from each question’s own table. As a support agent on tau2-bench it earned an average reward of 0.71 on airline and 0.80 on retail conversations.

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 (one command)

curl -fsSL https://raw.githubusercontent.com/chadstthomas-svg/terrabeam/master/install.sh | sh     # Windows: irm …/install.ps1 | iex
terrabeam run tablinum/Moon-Beam-8B-A1B-GGUF-Preview

terrabeam picks the right file for your GPU or CPU, downloads it with the matching llama.cpp build and the Tablinum library, and serves an OpenAI-style API at http://127.0.0.1:8080/v1. Every question gets the library's best passages in the exact layout Moon-beam was trained on. Add --memory long-term for persistent memory across conversations.

Use it from Open WebUI, LM Studio, Ollama, Claude Code, Cursor, Continue, Cline, opencode or any OpenAI client: setup for each app. Agents installing the model: read AGENTS.md.

Plain llama.cpp (no library; it answers from its own weights only):

llama-server --hf-repo tablinum/Moon-Beam-8B-A1B-GGUF-Preview --hf-file Moon-Beam-8B-A1B-Q4_K_M.gguf --jinja -c 32768 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0
# faster on GPU, with the MTP draft head:
llama-server -m Moon-Beam-8B-A1B-Q4_K_M.gguf -md Moon-Beam-8B-A1B-MTP-Q4_K_M.gguf -ngld 99 \
  --spec-type draft-mtp --spec-draft-n-max 2 --jinja -c 32768 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0

Needs llama.cpp b12078 or newer (Ling v3 / bailingmoe3 and draft-mtp). Thinking is on by default; turn it off per request with "chat_template_kwargs": {"enable_thinking": false}.

Files

Each quant uses its own importance matrix. Agreement and KLD are measured against this model's BF16 on 8,393 mixed chat, code and math positions and on WikiText. tablinum.json lists every file with its SHA256 (terrabeam verifies them).

How much does retrieval help?

With thinking on, retrieved passages raised SimpleQA from 30 to 601 correct answers out of 1,000. The same comparison on EngramBench went from 61 to 436 correct out of 503. Answers are graded blind with AA-Omniscience’s grader template — a deterministic rule pass first, then an LLM judge — with partial answers and abstentions counted separately from correct answers.

Read the figures as text

The untrained Ling base also reached 59.2% on SimpleQA with the same passages and thinking on, compared with 60.1% for round 1b on the card protocol. The base used IQ3 and earlier settings. The clearest gain in these runs comes from giving the model relevant passages. Graded counts and settings — the numbers above are the re-measured card-protocol runs; the November 4 file keeps the preliminary graded counts for comparison.

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

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.

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

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

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).