Skip to content

BigMoeOnEdge

v0.11.0 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

Published 8d LLM Frameworks
✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

android c++ edge-ai gemma gguf gpt-oss
+7 more
inference llama-cpp llm mixture-of-experts moe on-device-ai qwen

Summary

AI summary

Adds Qwen3.6-35B-A3B as a supported one‑tap model with streaming optimizations.

Changes in this release

Feature Low

Adds Qwen3.6-35B-A3B (qwen35moe) as a supported one-tap model.

Adds Qwen3.6-35B-A3B (qwen35moe) as a supported one-tap model.

Source: llm_adapter@2026-07-19

Confidence: high

Feature Low

Updates README to include qwen35moe in Supported models and adds Qwen3.6 benchmarks subsection.

Updates README to include qwen35moe in Supported models and adds Qwen3.6 benchmarks subsection.

Source: llm_adapter@2026-07-19

Confidence: high

Feature Low

Provides comparative performance table for mmap baseline vs streamed configurations.

Provides comparative performance table for mmap baseline vs streamed configurations.

Source: granite4.1:30b@2026-07-19-audit

Confidence: low

Performance Medium

Streaming with overlap and dense‑weights anon improves token throughput to 5.0 tok/s (cache hit 65%).

Streaming with overlap and dense‑weights anon improves token throughput to 5.0 tok/s (cache hit 65%).

Source: llm_adapter@2026-07-19

Confidence: high

Performance Medium

Turbo streaming (k=6) raises throughput to 5.8 tok/s with reduced memory usage (91 MiB, cache hit 68%).

Turbo streaming (k=6) raises throughput to 5.8 tok/s with reduced memory usage (91 MiB, cache hit 68%).

Source: llm_adapter@2026-07-19

Confidence: high

Full changelog

Adds Qwen3.6-35B-A3B (arch qwen35moe) as a supported, one-tap model — a hybrid attention/SSM
MoE (256 experts, top-8, 41 blocks) whose routed experts stream through the existing qwen3moe path
unchanged. No engine change: the architecture is already in the registry.

Added

  • In-app one-tap download entry for Qwen3.6-35B-A3B (Q4_K_M, ~22.3 GB).
  • README: qwen35moe in Supported models + a Qwen3.6 benchmarks subsection.

On-device (model ~2× RAM)

Streaming with --overlap --dense-weights anon runs it stably where the mmap baseline collapses:

| Config | tok/s | Flash/token | Cache hit |
|---|---:|---:|---:|
| mmap baseline (no streaming) | 0.1 | — | — |
| streamed, cache 3000 (lossless) | 5.0 | 144 MiB | 65% |
| streamed, cache 3000, k=6 (turbo) | 5.8 | 91 MiB | 68% |

Figures are an indicative 96-token run, not yet the full 256-token best-of protocol.

Full changelog: https://github.com/Helldez/BigMoeOnEdge/compare/v0.10.0...v0.11.0

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track BigMoeOnEdge

Get notified when new releases ship.

Sign up free

About BigMoeOnEdge

All releases →

Beta — feedback welcome: [email protected]