Ornith

Ornith 1.5 (ornith-ai, MIT, August 2026) is a sparse mixture-of-experts family for agentic coding and tool use. The 35B-A3B member: 34.7B total (plus a 1.9B MTP head shipped separately), ~3B active per token — 256 routed experts, 8 active, one shared per layer; hybrid attention (full attention on 10 of 40 layers, linear attention elsewhere); 262k context; text + image input; thinking trace returned separately. Trained with a self-improvement loop where the model proposes its own training tasks and learns from its own attempts. Positioned against Qwen3.6-35B-A3B (claims significant wins across coding/agentic benchmarks) and above dense Gemma 4-31B on agentic coding.

AtomicChat (atomic.chat) is the local-app vendor whose GGUF builds of Ornith 1.5 are the carefully-measured ones: importance matrix computed on BF16 weights over ~5M tokens fanned across 15 GPUs, per-file KLD and top-1 agreement published for every quant, calibration corpora and raw eval logs public (AtomicChat/calib-corpora, -GGUF-metrics repos). The AD- prefix marks their builds. Their model card’s memory table is the canonical answer to “which quant for my VRAM”, including the MoE-specific advice: with a smaller card, keep experts in system RAM (--cpu-moe) rather than dropping quant rungs, because experts are 93% of weights but only 3B run per token — arithmetic stays cheap while bandwidth moves to RAM.

On the doombox: AtomicChat/Ornith-1.5-35B-A3B-AD-Q4_K-IQ4_XS.gguf (20.1 GB class, 4 llama-swap entries incl. a -ngram speculative variant) — the Q4_K attention + IQ4_XS experts split-precision build.

Links: Qwen model family (the stack Ornith builds on and competes with — architecture string qwen35moe) · Gestalt Labs (whose Ornstein name echoes it) · DFlash2 (the other speculative-decoding track; Ornith ships an MTP head)