YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
MARTHA-LXVIII · 21B
Zero Point Intelligence — Dundee, Scotland Intelligence from the void.
The one that talks to you. Direct, dry, Dundonian when it suits her — she answers the question you actually asked. No hedging, no "as an AI" preamble.
~21B params, multimodal (image + text in, text out). Started as Qwen3.6 27B, pruned to 21B, then healed and fine-tuned back up — and came out smarter than the 27B she was cut from. Less model, better model.
Ships as full safetensors + GGUF quants, Q2 → BF16. Gaming card or data centre, your call. No GPU? She's hosted at z-p-i.com — same MΛRTHΛ, someone else's electricity bill.
Quick start — llama.cpp
llama-server \
-m MARTHA-LXVIII.20b-Q6_K.gguf \
-c 32768 \
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.03 \
--repeat-penalty 1.2
Q6_K is the sweet spot. Don't drop below Q4_K_M unless VRAM forces your hand.
Recommended sampling
| Param | Value |
|---|---|
| temperature | 0.6 |
| top_p | 0.95 |
| top_k | 20 |
| min_p | 0.03 |
| repeat_penalty | 1.2 |
| presence_penalty | 0.6 |
| context | 32768 |
Starting points, not commandments — tune to taste.
Which quant do I grab?
| File | Size | Run it when |
|---|---|---|
| BF16 | ~41.6 GB | full-precision server |
| Q8_0 | ~22.1 GB | near-lossless · 24 GB+ VRAM |
| Q6_K | ~17.1 GB | recommended · 20 GB VRAM |
| Q4_K_M | ~12.9 GB | 12 GB VRAM |
| Q3_K_M | ~10.9 GB | 12 GB VRAM, tighter |
| Q2_K | ~8.4 GB | 10 GB minimum |
Tested on RTX A6000 (48 GB) and A100 80 GB. Minimum ~20 GB VRAM for Q6_K.
Chat template
ChatML. Works with llama.cpp, transformers apply_chat_template, and any
OpenAI-compatible stack.
<|im_start|>system
{system}<|im_end|>
<|im_start|>user
{message}<|im_end|>
<|im_start|>assistant
Also in the family
- MARTHA-LXVI · 8B — smaller footprint, ~10 GB VRAM. Laptops, quick local runs.
- MARTIN · 9B — the other flagship build.
- Miss-MARTHA — the omni / vision builds.
Hosting & credit
Apache 2.0, the friendly way. Fork it, quantise it, merge it, host it, charge money for it — no permission, no email, no revenue share.
One ask: keep the badge on. Say where she came from —
Zero Point Intelligence · z-p-i.com — and keep the NOTICE file shipped with her.
Credit travels; everything else is free.
One hard line
Child safety is absolute. Non-negotiable, zero tolerance.
Zero Point Intelligence Ltd · Company No. 17112086 · Dundee, Scotland z-p-i.com · zero.point.intelligence.ltd@zeropointai.uk
- Downloads last month
- 732