Gemma 4 31B (dense) — Core AI

Apple Core AI (.aimodel) conversion of Google's Gemma 4 31B dense text decoder, ported directly from the QAT release google/gemma-4-31B-it-qat-q4_0-unquantized. Decode-only, runs on the stock pipelined engine on Apple Silicon (Mac-class, ~16 GB).

Frontier dense, unblocked by a custom Metal kernel. Gemma 4 31B's full (global) attention layers have a 32-head × 512 Q tensor that overflows MPSGraph's GPU decode scratch heap — the stock SDPA crashes at the first token (apple/coreai-models#27, the same bug as the 12B). This bundle ships a custom flash-decode SDPA kernel on the full layers (block-GQA over the 31B's 4 global KV heads) that removes the offending op, so the model runs.

Use it

▶️ Run it (source) — the ChatDemo runner (GUI + CLI, one app for every chat model in the catalog):

git clone https://github.com/john-rocky/coreai-kit
open coreai-kit/Examples/ChatDemo/ChatDemo.xcodeproj
# → Run, then pick "Gemma 4 31B" in the model picker

# agents / headless (macOS):
cd coreai-kit/Examples/ChatDemo
swift run chat-cli --model gemma-4-31b --prompt "What can you do, offline?"

💻 Build with it — complete; the glue is kit API, copy-paste runs:

import CoreAIKit

let chat = try await ChatSession(catalog: "gemma-4-31b")
let reply = try await chat.respond(to: prompt)
// reply: the answer, generated fully on-device

The take-home is Examples/ChatDemo/Sources/QuickStart.swift — this exact code as one typed function, no UI; the CLI is an argument shell over it, and the GUI drives the same ChatSession across turns for its transcript. Multi-turn? Hold the ChatSession and call respond(to:) per turn — it keeps the conversation history; streamResponse(to:) yields tokens as they decode.

Integration checklist

SPM: https://github.com/john-rocky/coreai-kit → product CoreAIKit
Info.plist: none needed
Entitlements: none needed (macOS)
First run downloads the model — 18.0 GB (Mac) — then it loads from the local cache (Application Support; progress via the downloadProgress callback)
Measure in Release — Debug is ~3× slower on per-token host work

Bundle (`gpu-pipelined/`)

bundle	quant	size	decode (M4 Max)
`gemma4_31b_qat_decode_int4linsym_msdpa_g8`	int4 (q4_0-aligned absmax)	19 GB	17.2 tok/s (prefill 22.1)

int4 from Google's QAT checkpoint (q4_0 grid). A frontier 31B at int4 is bandwidth-bound, so decode is in the MLX-parity range — the value is "Core AI runs a frontier dense model the stock engine cannot." Mac-only (exceeds the iPhone memory budget). The _g8 suffix is the higher-occupancy flash-decode kernel (8 SIMD-groups per head split the global layers' KV scan; same numerics).

Architecture

Clean dense gemma4 text decoder — no PLE / AltUp / Laurel / MoE / KV-sharing. 60 layers, hidden 5376, 32 heads, vocab 262144, softcap 30, tied embeddings. 5:1 sliding:full; dual head_dim (sliding 256 / full global_head_dim 512); full layers use num_global_key_value_heads 4 with attention_k_eq_v (value = raw k_proj). Both attention shapes ride one growing KV pair, so the bundle loads on the stock CoreAIPipelinedEngine (2 states, no engine patch); the full layers' SDPA runs as a custom Metal flash-decode kernel.

Usage

huggingface-cli download mlboydaisuke/Gemma-4-31B-CoreAI \
    --include "gpu-pipelined/gemma4_31b_qat_decode_int4linsym_msdpa_g8/*" \
    --local-dir ./gemma4-31b-coreai

COREAI_CHUNK_THRESHOLD=1 llm-runner \
    --model ./gemma4-31b-coreai/gpu-pipelined/gemma4_31b_qat_decode_int4linsym_msdpa_g8 \
    --prompt "What is the capital of France?" --max-tokens 64 --chunk-size 1