One-line: A 47.7B MoE (8 experts, top-2, ~30% active) upcycled from a dense 14.8B core, refined over two SFT epochs on a 99,661-example code+math curriculum, MIT-licensed. ⚠ weak coding numbers for the size (HumanEval 70.12%, MATH-500 38.4%) and zero agentic/tool-use evals — agent-loop suitability unknown. ⚠ custom architecture (PinkElephantForCausalLM, auto_map custom code) — stock llama.cpp/transformers won’t load it without the repo’s custom model code.
Coding benchmarks
Benchmark
Score
Source
HumanEval
70.12
lab-reported (115/164 pass@1, greedy; parent 14B was 71.34)
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
—
not published
Aider polyglot
—
not published
SWE-bench Verified
—
not published
SWE-bench Multilingual
—
not published
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
not published
τ-Bench
—
not published
ToolACE
—
not published
GAIA
—
not published
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
not published
GPQA Diamond
—
not published
MATH
38.4 (MATH-500)
lab-reported (parent 14B was 63.4 — the MoE upcycle actually regressed math)
AIME 2025
—
not published
Efficiency
Active params: ~14.66B (30.7%) of 47.7B total — MoE, 8 experts, top-2 routing
Context length: 16,384 tokens
VRAM (fp16 rough): ~95 GB full weights (all experts must be resident); ~29 GB if only active params streamed — practical local use needs quantization + custom loader
Architecture note: custom PinkElephantForCausalLM with auto_map custom code — will NOT load in stock llama.cpp/transformers; requires the repo’s modeling code.
What makes it notable
A fully-public research lineage (dense 14.8B → upcycled 48B MoE → refined 48B-S) with every artifact MIT-licensed and auditable eval JSONs in-repo. But for the coding/agent use case it’s not competitive: HumanEval 70.12% is well below what current 7B-14B models achieve, MATH-500 regressed vs its own dense parent (38.4 vs 63.4), and there are no agentic evals at all. The custom architecture is a deployment blocker for the standard local stack. Tracked because it qualified on structure (MIT, text-gen, lastModified today, 1.4k downloads); treat as a research curiosity, not a coding workhorse.