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 pages a model's experts from SSD, so a laptop runs models that would never fit in its memory. Pick your machine — see what runs, and what it costs.

3 of 3 tested models run on a 16 GB Mac · reserving ~4 GB for the OS

Qwen3-30B-A3B
fine-grained (128e/top8)
fidelity 97.4% · ppl +0.5452%
Runs · Strata~21.9 tok/s

Too big for RAM — Strata streams experts from SSD, keeping 51% hot.

your 12 GB usable
Full model17.2 GB
With Strata9.2 GB working set
GLM-4.5-Air
flat-router mid (glm4_moe)
fidelity 97.8% · ppl -0.4917%
Runs · Strata~2.8 tok/s

Too big for RAM — Strata streams experts from SSD, keeping 9% hot.

your 12 GB usable
Full model60.1 GB
With Strata8.9 GB working set
DeepSeek-V2-Lite
small MoE (deepseek_v2)
fidelity 100% · ppl +0%
Fits in RAM~49.3 tok/s

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

your 12 GB usable
Full model8.8 GB
With Strata8.8 GB resident
Strata working set (fits your RAM)full model (won't fit resident)tok/s measured on our M5 Max — verify yours

More classes landing next (paged-runtime support in progress): Mixtral-8x7B · gpt-oss-20b · Qwen3-Next-80B-A3B

$ pip install strata
$ strata bench --model mlx-community/Qwen3-30B-A3B-4bit
→ real tok/s + fidelity + RAM, measured 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.