Ornith 1.5 35B-A3B (MoE) vs Muse Glimmer 30B

Analysis

Ornith 1.5 35B-A3B and Muse Glimmer 30B are two of the closest peer launches tracked on this site: both are local coding-agent releases from the same August 2026 window (Muse Glimmer on 10 August from Meta, Ornith 1.5 on 19 August from Ornith AI), and Ornith AI's own launch materials name Muse Glimmer directly as one of the dense models the 35B-A3B benchmarks itself against. What makes the comparison worth more than a spec-sheet skim is that the two solve the local-inference-cost problem in almost opposite ways: Muse Glimmer stays fully dense and shrinks its KV cache with a sliding-window attention cap, while Ornith 1.5 routes through a mixture of experts so only a fraction of its parameters ever do compute on a given token.

Ornith 1.5 35B-A3B is continues-pretrained on Alibaba's Qwen3.5 checkpoint, and it shows in the architecture: 40 layers built from ten repeats of three Gated DeltaNet layers followed by one Gated Attention layer, so only 10 of 40 layers keep a KV cache at all (2 KV heads x 256 head dim), and MoE routing means only about 3 of its 35 billion parameters are active on any given token. Muse Glimmer 30B takes a different route to the same goal: every one of its 52 layers is real attention, but 39 of them (a repeating 3-local/1-global pattern) cap their cache to a 2,048-token sliding window, on a narrow 32-query/2-KV-head ratio, and because it's dense, all 27.8B of its parameters are active on every token regardless. At 131,072 tokens, the context length where Muse Glimmer's native window ends (Ornith continues on to 262,144, inherited from Qwen3.5), Ornith 1.5 35B-A3B's KV cache is 2.7 GB against Muse Glimmer's 1.8 GB, close despite the very different mechanisms, since both keep the vast majority of their layers out of the cache entirely. Add each model's Q4_K_M weights and the full picture is 26.9 GB total for Ornith 1.5 35B-A3B against 21.0 GB for Muse Glimmer 30B; at a more typical 8,192-token context it's 24.1 GB against 19.2 GB, Ornith's larger 35B weight floor (21.3 GB) outweighing Muse Glimmer's 30B-class weights (16.9 GB) at either length. On benchmarks both report, already tracked independently in this site's own data rather than freehand-quoted from either vendor's table, Ornith 1.5 35B-A3B leads on every shared figure: GPQA Diamond 89.2 vs 83.5, SWE-bench Verified 79.0 vs 76.0, SWE-bench Pro 59.6 vs 51.2, and Terminal-Bench 2.1 67.8 vs 51.7, all despite activating roughly a tenth the parameters per token that Muse Glimmer does. Modality is a real point of difference too, though a murkier one than the benchmark table suggests: Muse Glimmer documents image input through a separate ViT-G/14 encoder shipped as its own mmproj file, while Ornith 1.5 35B-A3B's checkpoint carries Qwen3.5's vision and video tower in its config and ships a matching mmproj GGUF, but Ornith AI's own documentation never demonstrates or claims multimodal use, so treat that capability as present in the weights but unverified rather than a documented feature the way Muse Glimmer's is. Licensing favors Ornith slightly: both are unrestricted (MIT for Ornith, Apache 2.0 for Muse Glimmer), but MIT carries even less boilerplate. Tooling favors Ornith here: it has a native cross-platform ollama library tag (ornith-1.5:35b) working on NVIDIA, AMD, and Apple Silicon alike, plus GGUF, MLX, FP8, and NVFP4 builds already up from ornith-ai directly and community requants from bartowski, while Muse Glimmer's only day-one Ollama tag is Apple Silicon (MLX) only, leaving NVIDIA and AMD users to run it through llama.cpp or vLLM against Unsloth's or Meta's own GGUF builds instead.

Bottom line: If raw agentic-coding benchmark score is what matters and a 32 GB card (or 32 GB+ unified memory) is available, Ornith 1.5 35B-A3B is the clear pick: it leads Muse Glimmer 30B on every benchmark both report, using a fraction of the active compute per token to get there. If a single 24 GB card is the hard ceiling, the calculus flips: Muse Glimmer 30B fits comfortably even at its full native 131,072-token context (21.0 GB total), while Ornith 1.5 35B-A3B's Q4_K_M weights alone already push past a 24 GB card's real ceiling at even a modest 8,192-token context (24.1 GB total). Treat Ornith AI's benchmark comparison as vendor-reported until an independent group reproduces it, the same caveat that applies to any same-day model launch, including Muse Glimmer's own numbers when they first shipped.