mlboydaisuke/Qwen3.6-35B-A3B-CoreAI

Qwen3.6-35B-A3B — Core AI (gather_qmm kernel, 2.1× faster)

Apple Core AI (.aimodel) conversion of Qwen/Qwen3.6-35B-A3B (text decoder): Qwen3.5's hybrid GatedDeltaNet + gated-attention body with a 256-expert top-8 sparse MoE (+ shared expert). 35B total / ~3B active per token.

Part of the community Core AI model zoo: https://github.com/john-rocky/coreai-model-zoo (full card: zoo/qwen3.6.md).

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 "Qwen3.6-35B-A3B (MoE)" in the model picker

# agents / headless (macOS):
cd coreai-kit/Examples/ChatDemo
swift run chat-cli --model qwen3.6-35b-a3b --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: "qwen3.6-35b-a3b")
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 — 35.0 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

The gather_qmm kernel — 30.9 → 64.9 tok/s (2.1×)

Apple's GatherMM composite gathers the routed experts then runs a dense matmul that reads all 256 experts' weights every token — over-read-bound at 30.9 tok/s. This bundle uses a custom coreai_torch.TorchMetalKernel that takes the routed indices as a kernel input and reads only the 8 routed experts' weight slabs (8/256), so decode runs at active-param (~3B) bandwidth: 64.9 tok/s, 2.1×.

Quality is clean and unchanged. The kernel reads the sym8 scheme = the same symmetric-linear int8 (per-K-block-32) recipe the standard int8 bundle uses, via a bit-exact gather: 0 introduced flips / 18 vs fp16 (the shipped GatherMM int8 was 14/16 vs the bf16 oracle; this matches it). So this is a pure speed win at the same quality.

bundle	size	decode tok/s	quality
gpu-pipelined/qwen3_6_35b_a3b_decode_sym8_gather/	35 GB	64.9	clean (0 flips/18 vs fp16) ✅

Mac-only (35 GB int8 is far past the iPhone limit; this is the 64/128 GB-Mac flagship).

Run

COREAI_CHUNK_THRESHOLD=1 llm-benchmark --model gpu-pipelined/qwen3_6_35b_a3b_decode_sym8_gather -p 128 -g 256 -n 3

The decode graph's input_ids is static [1,1]; prefill runs as S=1 pipelined steps. Convert your own with conversion/export_qwen3_6_moe_metal_decode_pipelined.py.

License

Apache-2.0 (upstream Qwen license). Conversion + gather_qmm kernel: community.