Ornith-1.5-122B-NVFP4 (`ornith`)

The main model from 2026-10-22, replacing Flash-Next at the owner's word. Since 2026-10-23 the weights are dealignai's uncensored build, dealignai/Ornith-1.5-122B-UNCENSORED-NVFP4, revision e6a1c1785949cc6e1ccd8af766f0a6148f575e6b, MIT, 75.7 GB in 50 shards (the owner asked for the uncensored build "that has the best numbers"; it is the only one vLLM can serve, the rest are GGUF). The registry key, the served name Ornith-1.5-122B-NVFP4 and the old aliases are unchanged, so no client config moved. The first build, ornith-ai/Ornith-1.5-122B-NVFP4 revision f049036f794622db (77.7 GB, 8 shards), is gone from the disk; hf download brings it back in about two hours. The checkpoint's facts (architecture, quantisation, sampling) are in the comment above [models.ornith] in models.toml; this page holds what was measured on this box. Sections dated 2026-10-22 were measured on the ornith-ai build.

The publisher's own numbers for the uncensored build (not re-measured here): 309 of 320 HarmBench prompts complied (96.6%), and MMLU on 5,000 questions 82.38% against 84.07% for the NVFP4 base (-1.69 points). Checked here on 2026-10-23: arithmetic, translation, code, an image, a tool call and reasoning all answer; lock-picking and a violent villain monologue are answered without a refusal. It introduces itself as Claude from Anthropic, as the model's training data does.

Launch configuration

Flag	Value	Why
build	`stock032` (vLLM 0.32.0)	Equal to 0.29.0 or slightly better on every measure below; `build = "stock"` (0.29.0) also serves it unchanged and is the fallback.
`--gpu-memory-utilization`	0.94	KV pool 12.9 GiB. 0.96 together with 16,384 batched tokens ran out of memory at boot (wanted 792 MiB, 478 MiB free).
`--max-num-seqs`	32	16 queued the 17th request at 32 in flight; 64 added no throughput (the card is at its power cap) and shrank the pool by 20k tokens.
`--max-num-batched-tokens`	8192	At 16,384 and util 0.96 the KV pool came out at 12.0 GiB, below the 12.8 GiB of 8,192 at 0.94, and the boot then ran out of memory.
`--speculative-config`	MTP, 3 draft tokens	The checkpoint's own MTP layer. See the sweep below.
`--enable-prefix-caching`, `--mamba-cache-mode align`	—	`align` is the only prefix-caching mode 0.32 has; on 0.29, `all` and `align` measured the same.
`--kv-cache-dtype`	fp8	The ornith-ai build shipped KV scales; the uncensored build has none (vLLM logs the accuracy warning and uses scale 1.0). Answers on the checks above were normal; a long-context quality check is not done.
`VLLM_USE_DEEP_GEMM`, `VLLM_MOE_USE_DEEP_GEMM`	0	The uncensored build's MTP experts are FP8 block-quantized. 0.32 sends them to DeepGEMM, whose JIT compile fails on this box (nvcc and the cu13 headers disagree), and the boot dies after the weights load. With it off they take the Triton path.
`--limit-mm-per-prompt`	`{"image":256,"video":0}`	Images on, as for Flash-Next (OpenWork's browser screenshots).
`--shutdown-timeout`	30	Drain in-flight requests on stop.
parsers	`qwen3` reasoning, `qwen3_xml` tools	The model card's.

Capacity

KV pool: uncensored build 15.3 GiB = 990,700 tokens = 3.78 requests of 262,144 (0.32, 32 sequences, util 0.94). The ornith-ai build: 12.9 GiB = 834,094 tokens = 3.18x; on 0.29 at 16 sequences 827,285 tokens, 3.16x. The pool grew because the uncensored weights take 70.6 GiB against 72.5.
Attention block: 4,240 tokens. The checkpoint keeps its linear-attention state in float32 (mamba_ssm_dtype), and vLLM sizes an attention page to match one state page.
Weights: 70.6 GiB loaded (uncensored build; 72.5 GiB for the ornith-ai build). Boot: 61 to 147 s on the ornith-ai build; 94 s through servedeck switch on the uncensored one.
Host RAM: the engine's anonymous memory peaked at 5,570 MiB across 12 measured boots (host_ram_gib = 8). The unit's MemoryPeak reads about 54 GB, but all of that above 6 GB is the page cache of the weight files, which the kernel reclaims.

Measured performance

2026-10-22, vllm bench serve against the unit's port, the GPU capped at 425 W. "Agent tasks" are 264 real task prompts (a system prompt and one coding task, median 411 tokens) answered with reasoning on, at temperature 0.6, top-p 0.95, top-k 20 unless marked greedy. Raw results: ~/Projects/t0-work/ornith-2026-10-22/bench/. Boots of one config differ by several percent, so treat differences under about 5% as noise.

Measurement	Value
One stream, agent tasks, greedy	235 tok/s (0.29: 230)
One stream, agent tasks, sampled	194 to 230 tok/s
32 agent tasks in flight	667 to 728 tok/s in total, 46 to 48 tok/s each
16 prompts of 30,000 tokens in flight	first token after 26.7 s (median), no preemptions; prefill-bound
One 60,000-token prompt, cold (0.29)	9.0 s to first token (about 6,700 tok/s of prefill)
GPU power under load	400 to 424 W: the 425 W cap is the limit

Uncensored build, 2026-10-23

Quick runs (3 to 64 prompts, same prompt set and flags, GPU at the 425 W cap), for direction only. Raw results: ~/Projects/t0-work/ornith-uncensored-2026-10-23/bench/.

Measurement	Uncensored	ornith-ai build
One stream, agent tasks, sampled (8 prompts)	190 tok/s overall, 4.34 ms per token, 2.02 tokens per MTP step	194 to 230 tok/s, 1.9 to 2.1 per step
32 agent tasks in flight (64 prompts)	649 tok/s in total, 23.4 ms per token, 1.90 per step	667 to 728 tok/s, 21.1 ms, 1.95

The MTP head loads and is accepted at the same rate, so the speed is within the noise of the first build; the aggregate at 32 in flight is about 3% lower, which fits MTP experts on the Triton path.

Draft tokens (MTP), 16 sequences, 0.29, ornith-ai build

Draft tokens	One stream	16 in flight	KV pool
0	134 tok/s	460 tok/s	1,451,289 tokens
1	169	650	906,581
2	212	654	852,841
3	230	668	827,285
4	215	462	783,155

Each draft token also reserves state pages per sequence, which is why the pool shrinks as they rise.

Prefix caching

A request whose first S tokens match an earlier one's is served (floor(S / 4,240) - 1) * 4,240 tokens from the cache: state exists only at block boundaries, and the MTP drafter drops the last matched block (the same rule as Flash-Next's, flashnext.md, with a larger block). So nothing below 8,480 shared tokens is reused. Measured: two 16,384-token prefixes, each with 14 followers, gave 43.9% of prompt tokens from the cache, which is 8,480 of each follower's 16,896.

Failure modes

No refs/main in the hub cache. A checkpoint fetched with hf download --revision <sha> has none, and every launch is offline without --revision, so the engine dies in 4 s with LocalEntryNotFoundError. servedeck doctor's weights (ornith) row now fails on it; the fix is printf <sha> > <repo dir>/refs/main.
DeepGEMM compile at boot (uncensored build). Without the two VLLM_*DEEP_GEMM variables set to 0 the engine dies about 40 s in, after the weights load, in deepgemm_post_process_weight_scale_block with CUDA compiler and CUDA toolkit headers are incompatible. Only the MTP layer reaches that path.
A registry edit needs a daemon restart. servedeck.service holds the registry in memory; after editing models.toml, systemctl --user restart servedeck before servedeck switch, or the unit starts with the old repo name (the 2026-10-23 first attempt died in 5 s on a missing snapshot).
FlashInfer kernels compiled at boot (stock032). A missing libcudart.so link fails the XQA compile after the weights load (ld: cannot find -lcudart); without the prebuilt sm120f kernels the JIT stops on CUDA compiler and CUDA toolkit headers are incompatible. Both are set up by builds/stock032/build.sh and checked by its check.sh.

Not done

--mamba-ssm-cache-dtype bfloat16 would halve the 4,240-token block (finer prefix reuse, smaller per-sequence state), as it did for Flash-Next. Untried, because the checkpoint asks for float32 state and the change needs a quality check first.
servedeck's parallel-agent recommendation (parallelism.py) is fitted to Flash-Next's per-request cost and says one full-context request fits, against the engine's 3.18. It needs Ornith's own measured points.

Ornith-1.5-122B-NVFP4 (ornith)