baa-ai/Ling-mini-2.0-RAM-11GB-MLX

Ling-mini-2.0 — 11GB (MLX)

Mixed-precision quantized version of inclusionAI/Ling-mini-2.0 optimised by baa.ai using a proprietary Black Sheep AI method.

This build is the RAM knee: the smallest budget the RAM probe predicts before quality falls away. Ling-mini-2.0 has 16B total and 1.4B active parameters per token.

Metrics

Metric	Value
Size	10.8 GB
MMLU (400 questions)	74.00%
MMLU vs BF16	97.0% of BF16
MMLU vs Uniform 4-bit	−1.25 points (within noise)
Document retrieval, Dash 8 Q300 maintenance manual (415 questions, right document in the top 3)	71.8%
Document Q&A, Boeing 787 manuals (150 questions, answer contains the expected value)	26.0%

The two document tasks were measured with the model as the reader of a retrieval-augmented console, against the same golden sets as five other readers (up to 90 GB): it was the fastest reader tested (median 1.9 s a question for retrieval, 3.5 s for a full answer) with retrieval within 1.5 points of the best, while long table lookups favour larger readers. MMLU is a 400-question set run with the model's chat template: up to 2048 tokens, then the model is asked for the letter if it has not given one. All rows in the table below were measured on that harness.

Model	Size	MMLU	MMLU % of BF16
BF16	32.5 GB	76.25%	100%
Uniform 4-bit	9.2 GB	75.25%	98.7%
RAM 11GB	10.8 GB	74.00%	97.0%

The repository includes the upstream modeling_bailing_moe_v2.py and configuration_bailing_moe_v2.py for transformers; MLX loads it natively with mlx-lm.

On this MMLU set the difference from uniform 4-bit is within noise (21 questions one way, 16 the other), and uniform 4-bit is 1.6 GB smaller.

Reasoning

Test	Score
GSM8K (100 questions)	95%
MMLU-Pro math (60 questions)	90%
Multi-hop chains, 4 to 6 hops (126), free answer up to 3072 tokens	57.9%
Multi-hop sums, 4 to 6 terms (90), free answer	97.8%
Multi-hop chains, terse one-word answer, 220 tokens, thinking off	3.2%
Multi-hop sums, terse answer, 220 tokens, thinking off	62.2%

Graded on the final answer only; an unfinished answer counts as wrong. GSM8K matches the final number as a number: Ling often ends with "#### <260>" or \boxed{} (a strict "#### 260" parser gives 53%).

Usage

from mlx_lm import load, generate



model, tokenizer = load("baa-ai/Ling-mini-2.0-RAM-11GB-MLX")

response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)

print(response)

---

Quantized by baa.ai

---

Black Sheep AI Products

Black Sheep AI builds private AI with the Bounded AI Architecture (BAA): three layers, each bound to something you control, on infrastructure you control. Nothing leaves your boundary.

Ram (Reasoning Layer) — Full-capability frontier models, shrunk to fit smaller GPUs with reasoning kept and certified. They rely strictly on your Data Modules for knowledge, so you swap the model as better ones ship.

Paddock (Knowledge Plane) — Binds your Data Modules directly to the reasoning layer: one per topic, period, person, permission or tenant. Permission, freeze, hand back or delete each one on its own.

Watchman (GRC Layer) — Governance, risk and compliance in a single, complete audit log: every module permission, every bind, every model's provenance. Five certificates before anything goes live.

Shepherd — Takes a frontier model, compresses it into a Ram reasoning model, certifies it with Watchman, binds it to your Paddock, and deploys the Bounded AI Architecture to your infrastructure: your cloud, a Mac Studio, or an air-gapped facility.

Learn more at baa.ai — Private AI for enterprises whose data cannot leave the building.