A 35B language model running on an iPhone using only 1–2.5 GB of peak memory

An open-source streaming MoE inference framework — SSD expert offload + Recover-LoRA + prerouter routing prediction.

English | 中文

edge0 is an open-source streaming MoE inference framework. It generalizes the production-proven recipe — SSD expert offload + Recover-LoRA + prerouter routing prediction — into an extensible framework. The backend is isolated by design: the current MLX backend runs on Apple Silicon, and additional platforms (CUDA, …) plug into the same core abstractions.

Two model tiers ship with the framework. Each tier is an end-to-end release: the released checkpoint, the trained LoRA adapters, and the trained prerouter heads work together as one unit.

Both checkpoints are built on open sparse-MoE base models (Qwen3.5-MoE 35B-A3B and the Ling 3.0 bailing hybrid respectively) and ship with the LoRA and prerouter training done for this framework — the adapter files are co-located with each checkpoint and load automatically, so edge0 serve <tier> runs the trained pipeline out of the box.

• OS / hardware : the MLX backend runs on macOS with Apple Silicon (M1/M2/M3/M4). The CUDA backend is on the roadmap — no other platforms are supported yet.

• Python : 3.10+ (3.12 recommended).

• Memory : ~2.9 GB peak active memory foredge0-35b , ~1.0 GB foredge0-8b (short contexts; seeBenchmark ). Add headroom for the OS, tokenizer, and long-context KV growth.

• Disk : the 4-bit checkpoints are ~23 GB (edge0-35b ) and ~4.2 GB (edge0-8b ); expert weights are mmapped and read on demand, they are not loaded into RAM up front.

• transformers-style usage :AutoModel /AutoConfig /AutoEngine resolve the tier from the model name;

• Backend isolation : all MLX code lives underedge0/backends/mlx/ ; the core logic (model specs, prerouter, streaming expert pool, server) depends only on the backend facade (edge0/backends/base.py ), so a new backend implements the same facade (backends/cuda/ is a reserved slot) with zero changes to core code;

• Adapters as safetensors : LoRA and prerouter weights are.safetensors files with provenance metadata (source, version, owner layers), resolved from the model directory orartifacts/ ;

• Model + adapters in one directory : a model directory holds both the base checkpoint (config.json /model*.safetensors / tokenizer) and that model's adapters; upgrading adapters swaps adapter files only — the base stays read-only and is never merged.

• SSD expert offload : expert weights are streamed from storage on demand; peak memory is bounded by the active set, not the parameter count.

• Prerouter : a trained head predicts expert routing one step ahead, so expert loads overlap the forward pass instead of stalling it —up to +59% decode throughput; the gain grows with storage latency, model size, and routed widthK .

• Recover-LoRA : the int4 base is frozen and LoRA adapters are trained by distillation from the FP teacher, recovering most of the quantization loss at 4-bit (seeQuality ). Adapters stay unmerged: one read-only base serves multiple adapter sets.

python3.12 -m venv .venv && .venv/bin/pip install -e '.[dev,fetch]'

The two tiers are published on Hugging Face — each repo bundles the base checkpoint and the trained LoRA + prerouter adapters in one directory, so a single download is a ready-to-run model:

• Edge0/Edge0-35B-A3B-preview (~23 GB)
• Edge0/Edge0-8B-A1B-preview (~4.2 GB)

.venv/bin/python scripts/fetch_models.py --tier edge0-35b --target-dir models
.venv/bin/python scripts/fetch_models.py --tier edge0-8b --target-dir models

.venv/bin/huggingface-cli download Edge0/Edge0-35B-A3B-preview     --local-dir models/edge0-35b
.venv/bin/huggingface-cli download Edge0/Edge0-8B-A1B-preview     --local-dir models/edge0-8b

Either way you end up with a directory like:

models/edge0-35b/
├── config.json, model-*.safetensors, tokenizer files   # base checkpoint
├── lora_edge0_35b.safetensors          # trained LoRA adapters
└── prerouter_edge0_35b.safetensors     # trained prerouter heads

Tier names resolve to local directories via environment variables (where you put the download is up to you):

export EDGE0_35B_MODEL=$PWD/models/edge0-35b
export EDGE0_8B_MODEL=$PWD/models/edge0-8b

Or skip the env vars entirely and pass the directory directly — the tier is auto-detected from the checkpoint's config.json:

edge0 demo models/edge0-35b
edge0 serve models/edge0-8b
edge0 demo edge0-35b

edge0 serve edge0-35b
curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Hello!"}],"max_tokens":32}'

edge0 chat edge0-35b --prompt "Explain streaming inference in one sentence."

python -m edge0 ... is equivalent to edge0 ....

from edge0 import AutoEngine
from edge0.server.chat import ChatMessage, ChatRequest, ChatSession

engine = AutoEngine.from_pretrained("/path/to/model")  # tier auto-detected
req = ChatRequest(
    model=engine.name,
    messages=[ChatMessage(role="user", content="Hello!")],
    max_tokens=64,
)
tokens, meta = ChatSession(engine, req).run()
print(engine._tok.decode(tokens))
engine.close()   # release mmaps / expert cache

examples/demo.py is the same minimal walkthrough (edge0 demo runs this exact path).

• Checkpoint : the original model directory (config.json ,model*.safetensors , tokenizer).edge0 serve <dir> /AutoEngine.from_pretrained(<dir>) detect the tier fromconfig.json .
• Adapters (LoRA + prerouter, safetensors) are resolved from either location automatically:
  • the model directory (recommended): side by side with the base, e.g. lora_edge0_35b.safetensors +prerouter_edge0_35b.safetensors ;
  • artifacts/ (repo root, gitignored): convert once from training-side npz exports viaedge0 convert-adapters --npz-dir ... .
• the model directory (recommended): side by side with the base, e.g.
• The published model repos bundle both the base checkpoint and the current default adapter release, so scripts/fetch_models.py produces a ready-to-run model directory. Check each model's doc page for its adapter provenance (training data, owner-layer layout).
• Both adapters are required for the prerouter + LoRA pipeline; if a file is missing, edge0 fails with a clear message (or pass--no-prerouter /--no-lora to run the plain base model).

All benchmarks were run by us with OpenCompass under identical settings and parameters for both the edge0 models (int4 + trained adapters + prerouter routing) and the original fp16 base models. The loss of the edge0 pipeline is small: 3.9 points on average for edge0-35b, 2.8 for edge0-8b (MMLU-Pro is even above the base). Max 100:

Measured with examples/bench.py (3.3k-token prompt prefill → 10 sampled warmup steps → 200 timed sampled decode tokens, 2 runs per tier):

Cold = first request after process start (expert weights fault in from SSD); warm = subsequent requests (page cache resident). Prefill numbers are throughput over a ~3.3k-token prompt (BENCH_LONG=1).

*Peak active memory at short contexts (MLX allocator peak; expert weights stream from SSD via mmap and are not resident). Long contexts add KV cache: ~3.3 GiB on edge0-8b at 3.3k tokens.

Reproduce:

python examples/bench.py edge0-35b    # via $EDGE0_35B_MODEL
python examples/bench.py edge0-8b    # via $EDGE0_8B_MODEL
pytest                 # unit tests (no real weights)
EDGE0_8B_MODEL=/path/to/edge0-8b pytest -m slow -q
.venv/bin/python scripts/e2e_smoke.py \
  --qwen-dir /path/to/edge0-35b --ling-dir /path/to/edge0-8b
scripts/generate_example.py   # full-pipeline API example
examples/demo.py       # minimal API walkthrough

Apache-2.0, including vendored third-party code (see NOTICE).