---
type: "post"
id: "moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2"
slug: "moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2"
title: "MoEShift alpha runs Qwen 3.8 Flash Next on Apple Silicon by streaming MoE experts from SSD"
description: "MoEShift 0.1.0 Alpha 1 is an open-source runtime that runs Qwen 3.8 Flash Next locally on Apple Silicon while loading mixture-of-experts (MoE) experts from SSD."
retrieval_nugget: "MoEShift 0.1.0 Alpha 1 is an open-source runtime that runs Qwen 3.8 Flash Next locally on Apple Silicon while loading mixture-of-experts (MoE) experts from SSD. Its measured comparison suggests a lower memory footprint than Strata in one long-prompt test, but this is an early release validated on a single Mac configuration, not a general performance guarantee."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: []
entities: ["MoEShift","Qwen 3.8 Flash Next","Apple Silicon","llama.cpp","llama.cpp-omni","Strata","ISTA-DASLab"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "MoEShift's alpha release adds a separate SSD-streamed Apple Silicon runtime for Qwen 3.8 Flash Next, built on llama.cpp; the priors covered Slotstream's similar approach."
prior_refs: ["memory:17a3e1b1-ba34-8990-ae23-a4a7bfd7b04e","memory:9281f7cf-a1bd-8e35-a1c7-c5988a05fbc9"]
source_url: "https://github.com/davide-troiani/moeshift/releases/tag/v0.1.0-alpha.1"
source_urls: ["https://github.com/davide-troiani/moeshift/releases/tag/v0.1.0-alpha.1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2/","markdown":"https://newruntime.com/posts/moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2.md","json":"https://newruntime.com/posts/moeshift-alpha-runs-qwen-3-8-flash-next-54f7c2.json"}
---

# MoEShift alpha runs Qwen 3.8 Flash Next on Apple Silicon by streaming MoE experts from SSD

## Retrieval answer

MoEShift 0.1.0 Alpha 1 is an open-source runtime that runs Qwen 3.8 Flash Next locally on Apple Silicon while loading mixture-of-experts (MoE) experts from SSD. Its measured comparison suggests a lower memory footprint than Strata in one long-prompt test, but this is an early release validated on a single Mac configuration, not a general performance guarantee.

## Summary
MoEShift 0.1.0 Alpha 1 is an open-source runtime that runs Qwen 3.8 Flash Next locally on Apple Silicon while loading mixture-of-experts (MoE) experts from SSD. Its measured comparison suggests a lower memory footprint than Strata in one long-prompt test, but this is an early release validated on a single Mac configuration, not a general performance guarantee.

## What happened
On October 30, davide-troiani released MoEShift v0.1.0-alpha.1, the first alpha of a runtime for local inference on Apple Silicon. Its reference model is Qwen 3.8 Flash Next using ISTA-DASLab’s GSQ-RCO IQ3_XXS quantization. Rather than keeping all model data in memory, MoEShift loads experts from an SSD. The release says it is built on llama.cpp through llama.cpp-omni and its Lookahead Expert Scheduling (LES) foundations.

The alpha includes an OpenAI-compatible HTTP server bound to localhost, intended to let compatible clients send requests to a local instance. It overlaps expert loading with Metal computation, and includes a demand-LRFU expert cache as well as separate budgets for prompt processing and generation. The release also lists growing context buffers and key-only indexer storage. Utilities extract and verify expert files from the original GGUF model, and the application creates an initial configuration automatically. These are implementation details and shipped components; the release does not establish that every compatible client or workload will work without changes.

MoEShift’s stated minimum requirements are an Apple Silicon Mac with at least 10 GB of RAM, a fast SSD, and at least 120 GB of free SSD space for the reference model and prepared files. The original model files must be downloaded separately. Thus, moving expert data off the machine’s working memory does not eliminate the storage requirement or make setup independent of the model files. The release points readers to the README for setup instructions and the architecture document for runtime details.

The release reports a long-prompt comparison against Strata on a Mac mini M4 Pro with 24 GB of unified memory. It gives inference times of 63.3 seconds for MoEShift and 129.9 seconds for Strata, and peak process memory of 7.5 GB and 14 GB respectively. The release points to complete results and measurement conditions in the README; the figures should be read as a reported test, not a universal benchmark. MoEShift is explicitly marked pre-release. The author says it was validated on M4 Pro and warns that performance varies with hardware, prompt length, and cache state. Users are directed to the repository’s issue forms to report installation or inference problems.

## What the archive adds
This is a new release in an existing local-inference thread, rather than a new general capability for all Macs. [Slotstream](https://newruntime.com/posts/slotstream-runs-a-125b-moe-model-on-16-1dfd4c/) covers a Mac-specific SSD-streaming approach and a claim to run a 125B MoE model on 16–64 GB Apple Silicon Macs. [Kapitsa v0.2.15](https://newruntime.com/posts/kapitsa-v0-2-15-an-open-source-apple-e8f061/) and [TensorFold 1.2.0](https://newruntime.com/posts/tensorfold-1-2-0-local-inference-engine-speeds-03faa0/) document related Apple Silicon inference work, including MoE support or performance. MoEShift adds its own alpha implementation and a specific comparison against Strata.

## Trend signal
MoEShift fits the broader pattern of local inference systems adapting execution and memory use to the hardware available, but none of the supplied SITE_PATTERNS directly describes SSD-streamed model inference. This release alone does not demonstrate a general shift in costs or production practice, so no pattern link is warranted.

## Why now
The release presents MoEShift as an alpha, with a reference model and a concrete Apple Silicon test configuration. That makes this a useful moment to examine the implementation and its constraints, but the source does not explain why the project was released on this date.

## Connections to other events
An earlier Telegram post described local inference as no longer a binary distinction between fitting a model and having a usable setup, citing quantization, MoE, and runtimes as ways to lower hardware thresholds. MoEShift is a concrete example of that approach: it moves expert loading to SSD, but adds substantial storage needs and remains sensitive to hardware, prompt length, and cache state. The earlier post’s broader point is not evidence that this alpha is ready for routine workloads.

## Why it matters
For builders with Apple Silicon, MoEShift offers a testable route to running a particular MoE model locally without keeping all of its expert data resident in memory. Its localhost OpenAI-compatible server also gives teams a familiar integration point to try, rather than requiring an entirely new request interface. The stated trade-off is practical: at least 120 GB of free SSD space, separately downloaded model files, and performance that depends on machine and workload.

The reported memory result may make experimentation more feasible on a 24 GB Mac, but it does not establish that the runtime is faster for every request. In the reported long-prompt test, MoEShift used less time as well as less peak process memory than Strata; other prompt lengths or cache states may change that result. Teams should treat this as a candidate for controlled local evaluation, not as a replacement recommendation. The alpha label also matters: installation friction and inference issues are expected to be part of the evaluation.

## What happens next
- **The project will document additional installation or inference issues as users test the alpha.** (3 months, 70%). Watch for: new release notes or repository issue discussions describing reproducible fixes or compatibility limits.

## Limits and uncertainty
The supplied release establishes a first alpha, stated requirements, listed components, and one comparison on an M4 Pro Mac mini. It does not provide the full measurement conditions in the release excerpt, independent replication, results across Mac models, or a broad comparison across prompts and cache states. The source explicitly says performance varies with hardware, prompt length, and cache state. It also does not establish production readiness, stability, or compatibility with any particular client beyond describing an OpenAI-compatible localhost server. The model files are downloaded separately, and at least 120 GB of free SSD space is required for the reference model and prepared files.

## Skill to build now
If you have an Apple Silicon Mac and enough SSD capacity, use the README’s setup instructions to reproduce the reference configuration before drawing conclusions from the headline comparison. Record the Mac model, memory, prompt length, cache state, inference time, and peak process memory for your own workload; report reproducible installation or inference problems through the repository’s issue forms. Keep the original model files and prepared files in the test plan, since the release requires them separately and needs substantial free SSD space.

## Reader takeaway
MoEShift’s alpha shows how SSD offloading can make local MoE inference a more plausible experiment on an Apple Silicon Mac with limited memory. The trade is not free: the reference setup requires at least 120 GB of SSD space, performance depends on workload and cache state, and the comparison is an early result from one M4 Pro test. Treat it as something to reproduce, not a general performance promise.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)