Ornith 1.5 122B (MoE) vs Qwen3.8-Flash-Next

Analysis

Ornith 1.5 122B and Qwen3.8-Flash-Next are close peers: both are roughly 120B-class mixture-of-experts releases, and Ornith AI's own model card benchmarks the first directly against the second. They share a hybrid layout (48 layers, only 12 of them keeping a KV cache) and both read text, images and video. What separates them is the memory bill, the license, and which kind of agent work you care about. Ornith 1.5 122B appeared on the Hugging Face hub on 16 October 2026 under the MIT license. Qwen3.8-Flash-Next arrived on 26 August 2026 as an experimental preview under the Qwen Community license; the production Qwen3.8-Flash is a separate hosted product.

The attention layouts match, so context length never separates these two. Both repeat three Gated DeltaNet layers and then one attention layer twelve times, with 2 KV heads of width 256 on the layers that cache, which puts the KV cache at 6.4 GB across the full 262,144-token window for each (0.2 GB at 8k). Qwen3.8-Flash-Next's attention layer is Qwen Sparse Attention, which cuts attention compute at long context but, per Qwen's tech report, does not shrink what has to stay in memory. The comparison therefore comes down to the weights. Qwen3.8-Flash-Next is 180B on disk, not the 125B in its headline: a 125B backbone, a 4B multi-token-prediction module and a 51B n-gram embedding table. This site prices all of it as VRAM, so its Q4_K_M weights are 109.6 GB and its total at 8k context is 123.0 GB, against 75.9 GB and 85.2 GB for Ornith 1.5 122B, about 31% less weight. That is a worst case. Qwen's tech report designs the n-gram table to be prefetched from host RAM, and vLLM and SGLang both documented that offload on release day. Counting only the backbone and the prediction module (129B), Q4_K_M would be about 78.6 GB of weights and 88.2 GB in total, within a few GB of Ornith's, so on a server stack with host RAM to spare the gap mostly closes. On hardware people actually own, the difference is sharper. A 96 GB RTX Pro 6000 holds Ornith 1.5 122B at Q4_K_M at 8k context, while Qwen3.8-Flash-Next's Q4_K_M does not fit it at all: this site's best fit there is Q2_K at 77.0 GB, below the Q4 floor it recommends. On a 256 GB M3 Ultra both run at Q4_K_M, and Qwen3.8-Flash-Next's 6B active parameters decode faster than Ornith's 10B: 60.8 tok/s against 49.6 in this site's estimates. On quality, Ornith AI's table has its model well ahead on SWE-bench Pro, slightly ahead on SWE-bench Verified and most agent suites, and behind on Terminal-Bench 2.1 and Toolathlon-Verified (see the chart below). Treat the widest gap with the most caution. The benchmark scores table further down shows Qwen3.8-Flash-Next's SWE-bench Pro as 62.5, which is Qwen's own README figure and the one this site stores, while Ornith's card lists it at 43.9. The card says its SWE-bench runs use the OpenHands harness with an anti-reward-hacking filter, so the two numbers are not the same measurement, and the card does not say whether Ornith AI re-ran Qwen3.8-Flash-Next or took its published scores. Qwen's README also reports GPQA Diamond 91.7 and LiveCodeBench 91.9, for which Ornith AI publishes no figure, so neither model can be called better overall from vendor numbers alone.

Bottom line: Choose Ornith 1.5 122B if your hardware is a single 96 GB card or a 128 GB machine that Qwen3.8-Flash-Next's full weights would overflow, if you want an MIT license, or if repository-level bug fixing is the job: it is the smaller download on this site's math, and Ornith AI's table has it ahead on SWE-bench Pro and Verified. Choose Qwen3.8-Flash-Next if you have 256 GB-class memory, or a server with host RAM for its n-gram table, and you care about decode speed (6B active parameters against 10B) or terminal-driven agent work, where it leads Terminal-Bench 2.1 by 2.2 points on Ornith AI's own table. Accept that it is a preview, and that every benchmark here is a vendor claim: run both on your own repository before committing.