The Apache-2.0 mixture-of-experts model shipped in February 2026, but a July write-up documents the local-runtime fixes — KV-cache reuse and disk-backed context restore — that finally made a long, cache-heavy chat feel fast on a single high-memory Mac.
2026-07-13 · by Moe Ameen
Qwen3.5-122B-A10B, the flagship open-weight model in Alibaba's Qwen3.5 line, was released in February 2026 under the permissive Apache 2.0 license. It is a mixture-of-experts (MoE) design: about 122 billion total parameters but only roughly 10 billion active per token, routed through a set of 256 experts. It pairs a hybrid attention architecture (linear-attention "DeltaNet" layers interleaved with standard attention) with a very long native context of 262,144 tokens, and it is multimodal on input — it accepts text and images and returns text. On independent aggregate benchmarks it sits at the strong end of open-weight models.
The news peg is not the release — it is usability. Because only ~10B parameters are active per token and the weights can be quantized to a low bit width, the model fits inside the ~96GB unified memory of a high-end Apple Silicon desktop (a Mac Studio M3 Ultra in the source write-up). But fitting in memory is not the same as feeling good to use. A detailed July 2026 write-up on mrzk.io documents the gap: on long, cache-heavy conversations, follow-up replies were taking minutes because the local inference stack was rebuilding the prompt cache every turn, dropping interrupted replies, and evicting the very checkpoints it needed to resume.
The fixes were in the tooling, not the model. The author's MLX-based runtime (a fork they call qMLX, built on rapid-mlx and tuned for Qwen's hybrid-attention layout) was patched to reuse the system-prompt KV cache instead of rebuilding it, to persist interrupted generations to history, and to restore the context cache from disk. The result they report is follow-up latency dropping from minutes to sub-second, with usable throughput even at 64K-token contexts. That is the shift worth noting: a frontier-class open model that was technically downloadable in February is now genuinely practical to run on one consumer machine. Treat exact figures — memory headroom, tokens per second, the extensible context ceiling — as one person's setup and confirm against the Qwen model card if they matter to you.
Here is the practical read for a creator: the cost and privacy of drafting text just fell to near zero, which means drafting text is no longer where anyone wins. When a capable model runs free on your own desk, the scarce thing becomes everything after the draft — a voice the audience recognizes, the media the words can't make on their own, and being on every feed at once. That is precisely the layer Kompozy is built to be. Run Qwen3.5-122B locally to reason over a long transcript and pull your ten best angles, then hand that text to Kompozy: the Persona Brief and banned-word filters rewrite it into your actual brand voice, and one idea fans into finished formats a raw model can't produce — Photo Posts, brand-exact Carousels through HyperFrames, Quote Graphics, Persona and HeyGen avatar video, Clipped Shorts, a Blog Article, and an Email Newsletter.
From there Kompozy does the half Qwen has no concept of: it reframes each output per platform, burns in branded captions, and schedules and publishes the whole package across nine social platforms plus email and blog from one queue, with Autopilot and a per-post review pipeline. And for the teams adopting local models specifically for cost or data control, Kompozy's bring-your-own-key support on the Founding tier keeps that discipline intact in the generation stack. Act on this news today by treating a local Qwen as your free drafting-and-reasoning brain and Kompozy as the engine that makes the output on-brand and ships it everywhere.
It is the flagship open-weight model in Alibaba's Qwen3.5 line, released in February 2026 under Apache 2.0. It has about 122 billion total parameters with only ~10 billion active per token (a mixture-of-experts design), a hybrid DeltaNet/attention architecture, a 262K-token native context, and multimodal input — it reads text and images and returns text.
Two things. Its MoE design keeps only ~10B parameters active per token, so with low-bit quantization it fits in the ~96GB unified memory of a high-end Mac. Then fixes in the local inference tooling — reusing the prompt cache, persisting replies, and restoring the context cache from disk — cut long-chat follow-up latency from minutes to sub-second. The model weights were unchanged; the runtime got the fixes.
Use a local Qwen as a free, private drafting and reasoning tool — mine long transcripts for hooks, react to screenshots — then bring the text into a content engine like Kompozy to rewrite it in your brand voice, generate the video/images/carousels the model can't make, and schedule and publish across platforms.