media Hugging Face Forums · just now · open_models

AiiStream Q3.6 enables running Qwen3.6-35B-A3B on 8-16 GB Macs

from English

The author has released AiiStream Q3.6, an open-source technique that allows the Qwen3.6-35B-A3B model to run on Apple Silicon devices with only 8 to 16 GB of RAM. By keeping routed experts on SSD and using early loading predictions, the method reduces MLX memory usage to approximately 1.7 GB while maintaining byte-identical outputs to the standard mlx_lm implementation.

Key performance metrics on a 16 GB M5 MacBook Air include:

15.0 tokens/s in controlled benchmarks, which is 72.6% faster than direct SSD streaming.
Approximately 11.8 tokens/s with other applications running.
A time-to-first-token of around 2.3 seconds for short prompts.

The technique is designed to work on 8 GB Macs as well, though performance figures for that configuration remain estimates based on the current release.

Importance 1/3 New harness with differentiators Trust 1/3 Hugging Face Forums Inference efficiency Local inference

Read original

github llama.cpp · 21h ago · 1 view

llama.cpp b11307 preserves batch order for speculative decoding

The llama.cpp project released build b11307, which fixes layer input ordering to preserve the original batch sequence during speculative decoding. This change ensures that token order is correctly maintained for unmasked NextN embeddings and compatible with tensor splits.

github llama.cpp · 1d ago · 1 view

llama.cpp b11301 fixes integer overflow in ggml/gguf

The llama.cpp project has released build b11301, which includes a fix for an integer overflow issue within the ggml and gguf components. This update addresses validation logic to prevent potential errors when handling zero-element tensors.

github llama.cpp · 1d ago · 3 views

llama.cpp b11302 fixes ThreadSanitizer data race in sparse indexer

The llama.cpp project released build b11302, which addresses a critical concurrency issue identified by ThreadSanitizer in the CI pipeline. The update resolves a data race within the sparse attention mechanism where multiple CPU threads were incorrectly writing to the same memory elements.

github llama.cpp · 1d ago · 2 views

llama.cpp b11298 adds dflash support to convert and feature extraction

The llama.cpp project has released build b11298, which introduces support for the dflash format. This update enables both conversion of models to this format and subsequent feature extraction capabilities.

github llama.cpp · 1d ago · 3 views

llama.cpp b11284 fixes OpenVINO GET_ROWS on quantized weight views

The llama.cpp project released build b11284, which includes a fix for the OpenVINO backend to correctly handle GET_ROWS operations on quantized weight views.