Strata

MODEL-ADAPTIVE OPTIMIZATION FOR SPARSE INFERENCE

Run sparse models faster, in less memory — without retraining.

Mixture-of-Experts models waste most of their serving cost moving expert weights around. Strata profiles how your model routes, calibrates competing policies under an explicit quality budget, and deploys the measured winner.

RUN BIG MODELS ON THE LAPTOP YOU HAVE

What can your Mac run?

Strata sizes each model's active working set to fit your memory budget, paging the rest of the experts from SSD — so a laptop can run a model far larger than its RAM. The working-set sizing is exact; the throughput below is an optimistic estimate until you verify it on your own device.

7 of 7 tested models fit a 16 GB working-memory budget · reserving ~4 GB for the OS

Peak memory is measured end-to-end (load + generate) under the documented benchmark config — it can vary with mlx-lm version, context length, and OS. Fidelity (paged vs resident output) is machine-independent. Throughput is device-dependent — measured on our 128 GB dev environment and optimistic; physical 16 GB constrained-device validation pending. Verify on your machine with strata bench.
Qwen3-30B-A3B
fine-grained (128e/top8)
fidelity near-exact*
Fits 16 GB budget~25.5tok/sest · 128 GB box

Won't fit in 16 GB resident (17.2 GB). Streamed — never assembles the full model — it loads and runs in a measured 12.2 GB peak (51% of experts hot, rest paged).

your 12 GB usable
Full model17.2 GB
With Strata12.2 GB measured peak
GLM-4.5-Air
flat-router mid (glm4_moe)
fidelity near-exact*
Fits 16 GB budget~2.9tok/sest · 128 GB box

Won't fit in 16 GB resident (60.1 GB). Streamed — never assembles the full model — it loads and runs in a measured 11.3 GB peak (9% of experts hot, rest paged).

your 12 GB usable
Full model60.1 GB
With Strata11.3 GB measured peak
DeepSeek-V2-Lite
small MoE (deepseek_v2)
fidelity bit-exact
Fits in RAM~81.4tok/sest · 128 GB box

Loads fully in memory — runs natively, Strata paging isn't needed.

your 12 GB usable
Full model8.8 GB
With Strata8.8 GB resident
Qwen3-Next-80B-A3B
ultra-fine (512e/top10)
fidelity near-exact*
Fits 16 GB budget~11.7tok/sest · 128 GB box

Won't fit in 16 GB resident (44.8 GB). Streamed — never assembles the full model — it loads and runs in a measured 11.8 GB peak (18% of experts hot, rest paged).

your 12 GB usable
Full model44.8 GB
With Strata11.8 GB measured peak
Mixtral-8x7B
coarse (8e/top2)
fidelity bit-exact
Fits 16 GB budget~2.0tok/sest · 128 GB box

Won't fit in 16 GB resident (26.3 GB). Streamed — never assembles the full model — it loads and runs in a measured 10.5 GB peak (25% of experts hot, rest paged).

your 12 GB usable
Full model26.3 GB
With Strata10.5 GB measured peak
Phi-3.5-MoE-instruct
16-expert (16e/top2)
fidelity bit-exact
Fits 16 GB budget~6.8tok/sest · 128 GB box

Won't fit in 16 GB resident (23.6 GB). Streamed — never assembles the full model — it loads and runs in a measured 10.2 GB peak (31% of experts hot, rest paged).

your 12 GB usable
Full model23.6 GB
With Strata10.2 GB measured peak
OLMoE-1B-7B-0125
tiny fine-grained (64e/top8)
fidelity bit-exact
Fits in RAM~162.5tok/sest · 128 GB box

Loads fully in memory — runs natively, Strata paging isn't needed.

your 12 GB usable
Full model3.9 GB
With Strata3.9 GB resident
Strata working set (fits your RAM)full model (won't fit resident)est · 128 GB box= optimistic tok/s from our dev machine (verify on yours)

bit-exact = paged output is bit-identical to the resident model (deterministic). near-exact* = quality-neutral ~1-ULP drift, not bit-identical (Qwen-family, GLM). Fidelity is machine-independent — hover any chip for the full measured metrics.

Benchmark config: peaks measured end-to-end (load + generate) on Apple M5 Max, mlx-lm 0.31.3, 140-token generation, per-model expert budget, greedy · commit 6562877. Peak can vary with mlx-lm version, context length, allocator, and OS.

More classes landing next (paged-runtime support in progress): gpt-oss-20b

$ pip install -e 'packages/strata[mlx]' # from the repo · PyPI soon
$ strata bench --model mlx-community/Qwen3-30B-A3B-4bit
→ real, device-measured tok/s + fidelity + RAM on your machine
102030402GB3.1GB4.1GB5.1GB6.1GBdecode tok/s · Qwen3-30B on a fixed memory budgetStrata-selected policystandard routing

Real measurement: Qwen3-30B-A3B, single node, physical cache-bypassed reads, 6 prompts × 2 runs per point. The gap is the product.

Measured, not promised

Every number carries its quality cost. When nothing beats standard routing within your budget, Strata says so.

GLM-4.5-Air
106B MoE
+76%
decode throughput at a fixed 21GB expert budget
quality within ±0.5% perplexity
Qwen3-30B
30B MoE
+64%
decode throughput at a fixed 5.1GB expert budget
quality within +0.5% perplexity
DeepSeek-V2-Lite
16B MoE
+101%
same-run baseline, 4GB expert budget
quality +0.5%, GSM8K gate passed

One command. Four stages. Zero guesswork.

01
Profile

Trace which experts your model's router actually activates, token by token, on your workload.

02
Calibrate

Sweep competing routing policies in normalized units and find each one's safe operating point under your quality budget.

03
Tournament

Race the calibrated policies on a real fixed-memory runtime with physical reads — the measured winner is the only one that counts.

04
Deploy

One config file, one OpenAI-compatible endpoint. Your application code doesn't change.

$ strata optimize --live --model your-org/your-moe --quality-budget 0.5
winner: cache @ 0.125σ, budget 32 → 19.78 tok/s (+101.4%)

No single policy wins everywhere. That's the point.

Across eight MoE architectures we measured, the best routing policy flips with the model, the memory budget, and the quality tolerance — including the best published method, which Strata implements as a baseline and beats or adopts per deployment. A hard-coded choice is provably wrong somewhere. A measured tournament never is.