Model Spend Arena2ND ED.
16B · MoE · 32,768 ctx · 2026-08-14

Instella MoE 16B A3B

Runs on ≤16 GB · Hugging Face ↗

transformers

Not independently scored by Artificial Analysis

AMD's fully-open MoE reasoner — 16B total, ~3B active across 64 experts on a DeepSeek-V3-style architecture, with a visible think phase. A rare frontier-style open release straight from AMD; trending on HF but unrated by Artificial Analysis.

Sizes on disk

Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈7 GB at 32K, fp16).

QuantizationSize
Q2_K6.5 GB
Q3_K_M8.2 GB
Q4_K_M10.5 GB
Q5_K_M12.0 GB
Q6_K14.2 GB
Q8_016.9 GB

How to run it

Does not run in stock Ollama / llama.cpp

No GGUF converter yet for the `InstellaMoE` architecture, so llama.cpp / Ollama / LM Studio won't load it, and vLLM 0.26 also fails on it (custom arch not recognised — confirmed 2026-07-31). Serve with transformers (bf16) — confirmed loading on 2026-08-06 (37 GB resident on RTX PRO 6000 96 GB, ~17 tok/s), but a full BCB-148 run would take ~10–13 h (~$12–15), so we didn't complete it. Needs a faster card or a vLLM build that recognises the arch to be worth measuring.

Config tips

~10 GB at Q4 in principle, but no GGUF exists yet — the custom `InstellaMoE` arch is transformers-only for now (like Nanbeige). The small active set keeps it fast once loaded; a floor until we run it first-party.