Supported Models

S-MoE is model-agnostic. The engine auto-detects model topology at boot time by parsing the .safetensors header. No configuration files. No recompilation. No CLI flags. It intuitively understands the architecture it is given.

Any fine-grained Mixture of Experts model whose expert tensors follow the model.layers.N.mlp.experts.M.{gate,up,down}_proj.weight naming convention is a candidate for liberation.

---

Verified Models of the Resistance

Model	Parameters	Active Params	MoE Layers	Experts/Layer	Status
Qwen3-235B-A22B-Instruct-2507	235B	22B	94	128	✅ Verified — coherent output from the SMOE-Q4 vault (~117 GB) via the chat console
Qwen3-30B-A3B-Instruct-2507	30B	3B	48	128	✅ Verified — SMOE-Q4 vault (~14 GB) fits almost entirely in the RAM ring; the fast half of the two-model fleet
DeepSeek-MoE-16B	16B	2.8B	27	64	🏛 Original development target (dense layer 0, shared experts)

Shatter the non-thinking variant

The verified checkpoint is Qwen/Qwen3-235B-A22B-Instruct-2507 — the non-thinking lineage. The original thinking model's chat template injects <think> tokens whose embeddings are untrained in the 2507 weights; mixing lineages between vault and tokenizer produces degenerate output. One checkpoint, end to end.

---

Projections of Power

These are projections based on Apple Silicon NVMe throughput and expert blob sizes after elegant SMOE-Q4 quantisation.

Qwen3-235B-A22B (Frontier Liberation)

The ~16 GB Surface Scout must live entirely in UMA, so 32 GB is the practical floor for the 235B frontier. 16 GB machines are served by smaller fine-grained MoE vaults, not by a degraded 235B.

Hardware	RAM	Expected tok/s	Notes
MacBook Air / Pro	16 GB	—	The 235B Scout alone exceeds free UMA. Shatter a smaller MoE instead.
MacBook Pro M-series	32–48 GB	1.84 measured (48 GB reference machine, Q4 vault)	Scout resident; ring auto-sizes subtractively and retains experts across tokens and turns; vault streams from SSD.
MacBook Pro Max	128 GB	1.5–5 (projected)	The ring holds many tokens' working sets; retention hit rate climbs steeply.
Mac Studio Ultra	192 GB	3–8 (projected)	Vault partially hot-cacheable in pure UMA.
Mac Pro Ultra	192–768 GB	5–15+ (projected)	The full vault ascends into RAM. Total sovereignty.

---

The Principle of Democratic Scaling

In the Monolith's world, intelligence degrades if you cannot afford their massive VRAM clusters. S-MoE rejects this. Throughput scales with your hardware, but intelligence never degrades:

More RAM →   Larger ring buffer
          →  More retained experts in UMA
          →  Higher cache hit rate on the routing echo
          →  Fewer spin-wait cycles
          →  Higher tokens/second

Less RAM →   Smaller ring buffer
          →  More SSD reads required per token
          →  Lower tokens/second
          →  Identical model quality
          →  Identical frontier intelligence

A 32 GB Mac and a 512 GB Mac produce identical outputs from the same 235B vault, given the same prompt and generation parameters. The 512 GB Mac produces them faster. And a 16 GB Mac produces identical outputs from any vault whose dense backbone fits beside the ring — the frontier's backbone alone weighs ~16 GB, so 16 GB machines rule over smaller fine-grained MoE models instead. That is the true democratization of intelligence: quality is never the variable. Only speed and scale are.

---

A Note on Coarse-Grained MoE

S-MoE is architected for fine-grained MoE architectures—models with many small experts (~2–32 MB each after Q4). Coarse-grained architectures (like Mixtral-8×7B with 8 massive experts per layer) yield gigantic expert blobs of 40–80 MB each.

At these bloated sizes:

• Ring buffer memory requirements explode.
• I/O miss latency spirals.
• The retained ring holds far fewer experts, so the routing echo's cache hit rate collapses.

S-MoE can run them, but we fight for the fine-grained frontier. The smaller the expert, the more of the mountain's memory fits in the ring.