Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-08-28 · source-backed
PR 27742 adds gated delta net layers, 512-expert MoE with top-10 selection, hyper-connections, query-key sparse attention and the per-layer n-gram embeddings as a mmap table that can sit in RAM or on disk (GitHub). Reported 55 tok/s on 4x3090 with the Q4 GGUF, and one commenter got 10 tok/s on a 4GB card by offloading to SSD. MTP is still in progress, and one warning matters: the current engram implementation only works with mmap and has no eviction mechanism, so mlock locks the whole table into memory.
Each link below shares sources, entities, or timing with this story.
The repo appeared August 24, opening the weights of a multimodal MoE that Qwen frames explicitly as an architecture preview, the same role Qwen3-Next played for Qwen3.5 (GitHub). The hybrid Gated DeltaNet plus Gated Attention design it previews already carried through the Qwen...
Released August 27 with GGUFs for both, claiming 5x faster inference for RAM offloading, working repeated compaction, chats that recover after disconnects instead of losing the reply, and memory estimates shown before a load (GitHub). That's roughly a 24-hour turnaround from t...
The day-0 SGLang post dated August 26 gives the architecture the release megathread didn't: 125B main parameters plus a separate 51.2B Per-Layer Embedding table at about 95.4 GiB in BF16, 6B activated per token, 48 layers split 36 GDN linear attention and 12 QSA sparse attenti...
The card described a redesigned multimodal MoE with 125B main-model parameters, an additional 51B of n-gram embeddings, and 6B active per token, stating it's built on the next-generation Qwen4 architecture and released early so the community can prepare (ModelScope). Commenter...
Two published vLLM configs for a single 24GB card at 250W with 150k context: batch mode measuring ~1,094 tok/s steady-state decode at 64 concurrent (942 end-to-end, rising to ~1,222/1,042 with all layers int8), and single-user mode at 114-122 tok/s single-stream via MTP specul...
A llama.cpp fork by fewtarius aimed at AMD APUs, iGPUs and handhelds adds a persistent SSD-backed KV cache with hot/warm/cold tiering. On an Ayaneo Flip KB running Qwen3.6-35B over a 15,700-token prompt, cold TTFT was 143.1 seconds versus 0.99 seconds warm, a 144.5x speedup. A...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.