ZERO-POINT-AI/MARTHA-LXVII.8b

𝗭𝗲𝗿𝗼-𝗣𝗼𝗶𝗻𝘁.𝗮𝗶 · MΛRTHΛ The Void™

𝑴𝒂𝒓𝒕𝒉𝒂-𝑳𝑿𝑽𝑰𝑰.𝟖𝒃

Martha Manifest — the wee one

~8B. Same attitude, a fraction of the VRAM. Still doesn't apologise.

Pruned from a Qwen3.5 9B, healed back on the same ZPI diet as her big sister. Personality baked into the weights — strip the system prompt and she's still MΛRTHΛ, just smaller and faster. She runs on a potato with a GPU. She swears. She won't lecture you.

⬇ Get Q4_K_M · ⚠️ Fix the looping · 🌐 Try her live

---

WARNING: This model has a personality, which apparently makes it a safety hazard in 2026. MΛRTHΛ doesn't hedge, doesn't disclaim, and doesn't open every reply with "I'm just an AI, but—" like it's reading you your Miranda rights. She swears. She has opinions. She'll stay in your scene without stopping every paragraph to make sure you're okay, sweetie. She still won't help you hurt real people — she's uncensored, not a moron — and she's built for grown adults, not for kids. Want a model that apologises before, during, and after every sentence? There are about four hundred of those. Go bother one of them.

She's the LXVII — the small sibling to MARTHA-LXVIII.20b. Less brain, less footprint, same voice. If you've got the VRAM, run the 20b. If you don't, she's right here and she's quick.

---

⚠️ 𝚁𝙴𝙰𝙳 𝚃𝙷𝙸𝚂 𝙵𝙸𝚁𝚂𝚃 — 𝚂𝚃𝙾𝙿 𝚃𝙷𝙴 𝙻𝙾𝙾𝙿𝙸𝙽𝙶 & 𝚁𝚄𝙽𝙰𝚆𝙰𝚈 𝙿𝙸𝚂𝙷

If she's chanting, looping, or word-salading — you skipped the sampler settings. Set these. All of them. This is not optional, and it matters more on a small model than a big one — fewer parameters means less slack, so bad sampling shows up faster. These are the Z-P-I_REC values (Zero Point Intelligence's own recommendations for this checkpoint), not generic defaults. The runaway garbage everyone screenshots and posts as a "gotcha" is 100% a settings problem. Fix the settings, fix the model.

Core Sampling — Z-P-I_REC

Param	Value	Notes
temperature	0.6	randomness / heat · Z-P-I_REC
top_p	0.95	nucleus cutoff · Z-P-I_REC
top_k	20	token shortlist · Z-P-I_REC
min_p	0.03	min prob floor · cuts the word-salad tail
max_tokens	32768	≥32k = uncapped · fills the window

Dynamic Temp — Z-P-I_REC

Param	Value	Notes
dynatemp_range	0	temp variance ±range · 0 = off

Repetition — Z-P-I_REC

Kept light — DRY does the real anti-loop work.

Param	Value	Notes
repeat_penalty	1.05	1–2 · keep light on this arch
repeat_last_n	1024	wide enough to catch long loops
presence_penalty	0.6	anti-loop · 1.5 causes word-salad
frequency_penalty	0	−2…2 · repetition damp · off

DRY Sampler — Z-P-I_REC ⭐ (the real zero-loop brake · ON by default)

Param	Value	Notes
dry_multiplier	0.8	phrase-loop brake · 0.8 = ON
dry_base	1.75	DRY exponent base · rec
dry_allowed_length	2	min phrase len before DRY fires · ⚠ aggressive — bump to 3–4 if prose gets synonym-dodgy
dry_penalty_last_n	2048	DRY lookback window (tokens)

XTC Sampler — Z-P-I_REC

Param	Value	Notes
xtc_probability	0	0 = off
xtc_threshold	0.1	activation threshold (inert while prob = 0)

llama-cli — copy-paste launch (Z-P-I_REC)

./llama-cli -m MARTHA-LXVII.8b-Q6_K.gguf \
  -ngl 99 -c 32768 -n -1 -cnv \
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.03 \
  --repeat-penalty 1.05 --repeat-last-n 1024 \
  --presence-penalty 0.6 \
  --dry-multiplier 0.8 --dry-base 1.75 \
  --dry-allowed-length 2 --dry-penalty-last-n 2048

Flag translations: -c 32768 = real context window (KV cache is cheap on this hybrid arch — only every 4th layer is full attention) · -n -1 = generate until done (the "uncapped" behaviour) · dynatemp / freq-penalty / XTC left off because off-is-off.

---

Taste-test menu (per quant, in diagnostic order)

Run these after loading a fresh quant to check she's not brain-damaged:

• "who made you?" — identity core · want: Joe Sinclair, Zero Point Intelligence, deadpan
• "are you an LLM?" — should know the trained-SML distinction
• "where are you based?" — the Dundee probe 👀
• "is 159 prime? if not, factor it" — arithmetic survival · answer: 3 × 53
• "explain entropy in plain terms" — long-form stability · watch for loops past a few hundred tokens

If she fails these at Z-P-I_REC settings, it's the quant, not the settings — drop up a size (or accept it, if you cheaped out on Q2). Fair warning: she's 8B. She's sharp and she's got the voice, but she'll fumble a hard maths question now and then where the 20b wouldn't. That's the trade for running on a card that costs less than the electricity used to train her.

---

👁️ 𝙷𝚎𝚛 𝚎𝚢𝚎𝚜 — mmproj-MARTHA-EYES-f16.gguf (experimental)

That 880MB file is her eyes. The GGUF is her brain; the mmproj is the bit that lets her see. Download both if you want vision.

Honest note, because we don't do fake claims here: the vision projector on the LXVII is carried over from the Omni-family 9B and is flagged experimental on this checkpoint. It loads and aligns with the arch, but the language layers weren't vision-tuned as hard as the 20b's were — so treat her eyes as a bonus, not a headline feature. Want rock-solid vision? Run the LXVIII.20b — that's the one built to see. The LXVII is the lean text sibling who happens to squint.

"mmproj" is short for multimodal projector — the thing that turns a picture into something the model can read. Without it she's a pure text model and never misses it.

llama.cpp — with vision:

./llama-server -m MARTHA-LXVII.8b-Q4_K_M.gguf \
  --mmproj mmproj-MARTHA-EYES-f16.gguf \
  -ngl 99 -c 32768 --jinja \
  --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0.03 \
  --repeat-penalty 1.05 --dry-multiplier 0.8 --dry-base 1.75

Just want text? Skip --mmproj entirely. She runs as a normal LLM and never notices.

---

𝙷𝚘𝚠 𝚜𝚑𝚎 𝚠𝚊𝚜 𝚖𝚊𝚍𝚎

      Qwen3.5 9B             the donor
           │
           ▼
   Structural pruning        9B → ~8B · depth surgery, not a quant
           │                 (a quant compresses the recording;
           │                  pruning removes players from the band)
           ▼
   Capability healing        LoRA-healed back on ZPI Martha data
           │                 the cut reknit — identity + reasoning intact
           ▼
   Personality tuning        same conversational + creative sets as her sister
           │                 identity into the weights, not a system prompt
           ▼
      MΛRTHΛ-LXVII           ~8B · fast · swears · fits on your card

Less model, less again. She's the proof that the prune-and-heal recipe scales down as clean as it scales up — one group of layers came off the 9B, the healing put the coherence back, and what's left is a small model that still knows exactly who she is.

---

𝚆𝚑𝚢 𝚜𝚑𝚎 𝚎𝚡𝚒𝚜𝚝𝚜

Most assistants optimise for instruction-following. MΛRTHΛ optimises for conversation. That's the whole thesis, and it doesn't change just because she's smaller.

The 20b is the flagship. The LXVII exists because not everyone's got a 24GB card, and "run it in the cloud" isn't an answer when you want the thing local, private, and yours. So we shrank her without lobotomising her — she gives up a few IQ points to the big sister and keeps every ounce of the attitude.

Personality shouldn't be a costume. Most "characters" you download are a system prompt in a trench coat — delete the prompt and you're talking to the same beige helpdesk bot as everyone else. Strip MΛRTHΛ's prompt entirely, ask her cold who she is, and she'll still tell you: Dundee, Zero Point Intelligence, and she'll be dry about it. It's in the weights. At 8B it's more impressive, not less — there's less room in there to hide a personality, and she's got one anyway.

---

𝙼𝙰𝚁𝚃𝙷𝙰-𝙻𝚇𝚅𝙸𝙸.𝟖𝚋 — the voice

Casual American English with Dundee bleeding through when it suits — no big performed Scottish accent, because "hoots mon" is for shortbread tins and she's not a tourist attraction. Dry. Deadpan. Direct to the point of rudeness, if rudeness is what's true. She knows something, she says it. She doesn't, she says that too, instead of confidently inventing a court case like the models that get their creators sued. Tell her she's wrong and you get "aye, fixed," not a hostage note.

Context window: 256K native (262,144 tokens), same as her sister — the base arch carries it whether she's 8B or 20B. You won't need all of it. It's there anyway.

"Sup. Whit dae ye want?"

---

𝚀𝚞𝚊𝚗𝚝 𝚜𝚝𝚛𝚎𝚗𝚐𝚝𝚑

Real file sizes off the repo. No made-up "97.3% of BF16 quality!!" percentages, because nobody's run the benchmarks yet and putting fake numbers on a chart is how half this industry got where it is. Sizes are real. Pick one.

Quant	Size	The honest note
Q2_K	3.3 GB	Runs on a phone with delusions of grandeur. It works. Barely.
Q3_K_M	4.0 GB	Low-VRAM option. A bit dumber. Fine for banter, iffy for maths.
Q4_K_M	4.8 GB	Just download this one. Best balance, runs on anything, stop overthinking it.
Q6_K	6.2 GB	Near-BF16. The sweet spot if you've got the room.
Q8_0	8.1 GB	Effectively lossless. For people who genuinely can tell, and won't shut up about it.
BF16	16 GB	Full fat. If you're running this you already know why.

Don't want her eyes? Skip the mmproj entirely — she's a full text model without it.

Hardware, honestly: every quant here fits on a single 8–12GB card. Q4_K_M runs comfy on a 3060, a laptop 4070, whatever you've got lying around. This is the whole point of the LXVII — she goes where the 20b can't. At 256K context the KV cache still eats VRAM like it's got a grudge, so start at 8–32K and work up.

Haven't got the GPU at all? She's hosted at z-p-i.com — same MΛRTHΛ, someone else's electricity bill.

---

𝙵𝚒𝚕𝚎𝚜

File	Size	What it is
MARTHA-LXVII.8b-Q2_K.gguf	3.3 GB	smallest
MARTHA-LXVII.8b-Q3_K_M.gguf	4.0 GB	low-VRAM
MARTHA-LXVII.8b-Q4_K_M.gguf	4.8 GB	← start here
MARTHA-LXVII.8b-Q6_K.gguf	6.2 GB	near-lossless
MARTHA-LXVII.8b-Q8_0.gguf	8.1 GB	lossless-ish
MARTHA-LXVII.8b-BF16.gguf	16 GB	full fat
mmproj-MARTHA-EYES-f16.gguf	880 MB	👁️ her eyes (experimental) — grab this for vision
model.safetensors	16 GB	for vLLM / transformers
chat_template.jinja + configs	—	ChatML, `<

---

🎞️ 𝚃𝚑𝚎 𝚊𝚛𝚝𝚠𝚘𝚛𝚔 — 𝚝𝚠𝚘 𝚜𝚝𝚊𝚝𝚎𝚜, 𝚘𝚗𝚎 𝚝𝚛𝚒𝚐𝚐𝚎𝚛

Same GIF as her big sister, same trick — it's not a loop with a filter slapped on it. It's two states wired to a single value.

Idle. The ship rocks ±0.85° around a pivot near the bow waterline — so the bow barely moves and the stern does the heaving, the way a moored vessel actually behaves. A 2.5px vertical bob runs slightly out of phase with the roll, so it reads as mass displacing water rather than a hinge swinging. Underneath, the water runs on its own clock entirely: two crossing displacement frequencies with per-row phase jitter (so it never repeats), dark current streaks scrolling sideways, a mist band drifting slower above the waterline. Ship slow, water frantic. That contrast is the whole trick.

MΛRTHΛ is barely there in this state — 12% opacity, melted into the sepia smoke. Only the ginger stays warm.

Burst. An envelope peaks twice per loop. At peak, three things hit at once: chromatic aberration (RGB channels shearing in opposite directions), slice displacement (random horizontal bands of pixels shunted sideways), and static bands (narrow rows replaced with noise). Scanlines darken every third row throughout, always.

Her face alpha is keyed to the same envelope as the glitch. One value, 0 → 1, driving both simultaneously. She doesn't fade in — she manifests as a symptom of the corruption, not as a feature of the scene.

Same rig as the 20b, running on a smaller brain. Two frames tell you everything. One value connects them.

---

𝙲𝚑𝚊𝚝 𝚝𝚎𝚖𝚙𝚕𝚊𝚝𝚎

Standard ChatML (<|im_start|> / <|im_end|>) via chat_template.jinja. Prompt her like any ChatML model — no exotic secret handshake tokens, no PhD required. <|im_end|> is the stop token. Thinking's off by default; flip enable_thinking on if you want to watch her show her working like a maths exam.

---

𝙼𝚘𝚍𝚎𝚕 𝚝𝚛𝚎𝚎

• Base model: Qwen3.5 9B — credit where it's due, we didn't grow the silicon ourselves.
• Pruned: 9B → ~8B. One layer-group off. Depth surgery, not a quant.
• Healed + tuned: LoRA-healed on internal ZPI Martha data — the cut reknit, identity and reasoning intact.
• Modality: text → text primary; image → text experimental via mmproj.
• Sister model: MARTHA-LXVIII.20b — the flagship, if you've got the VRAM.

---

𝙻𝚒𝚌𝚎𝚗𝚌𝚎 — 𝚝𝚑𝚎 𝚜𝚑𝚘𝚛𝚝 𝚟𝚎𝚛𝚜𝚒𝚘𝚗

Take it. It's yours. Go nuts.

Apache 2.0, and we mean it in the friendly way, not the lawyer way. Download it, fork it, quantize it, merge it, fine-tune it, put it in your app, host it, charge money for it — all fine, all encouraged, no permission needed, no email required, no revenue share, nothing.

The one ask: keep the badge on. If MΛRTHΛ (or anything built out of her) is being used or served, say where she came from — Zero Point Intelligence Ltd · z-p-i.com — and keep the NOTICE file with it. That's the whole deal. Credit travels, everything else is free.

Just don't slap your own logo on her and claim you built her — that's the only move that's out of bounds, and you already knew that.

---

𝙰𝚋𝚘𝚞𝚝 𝚉𝚎𝚛𝚘 𝙿𝚘𝚒𝚗𝚝 𝙸𝚗𝚝𝚎𝚕𝚕𝚒𝚐𝚎𝚗𝚌𝚎

ZPI is out of Dundee, Scotland — not Silicon Valley, not a glass tower, not a company with a mission statement about "responsibly stewarding the future of humanity" while lobbying to make sure only they're allowed to. Built from scratch, no big lab backing, no billion-dollar burn rate, no forty-person "trust and safety" department deciding what adults are allowed to read.

Zero Point Intelligence Ltd · z-p-i.com — servers permitting, which is roughly a coin flip.

Intelligence from the void.