A surgically-scoped MLP-only, weight-only NVFP4 quant:
Component
Precision
Why
MoE experts (all 256) + shared-expert MLP
NVFP4 (W4A16)
~90 % of the weights; where the size lives
Full attention (q/k/v/o)
BF16
preserved — FP4 attention costs reasoning
GatedDeltaNet / SSM (linear_attn.*)
BF16
recurrent path must stay full precision
Vision tower (model.visual.*)
BF16
multimodal kept intact
Router gates, embeddings, lm_head, norms
BF16
routing/IO precision-sensitive
Weight-only (W4A16): weights are NVFP4, activations stay BF16. This is the key choice — the well-known NVFP4 reasoning penalty comes from FP4 activations (W4A4), not the weights. Weight-only sidesteps it and, on Blackwell, dequantizes to BF16 in-register (also avoiding the sm_120/121 grouped-FP4-GEMM failure mode).
Validation (measured on the served NVFP4 model, vs the BF16 parent)
Zero coding-capability loss, zero re-censoring, fully coherent. Quantized on a B200 (Blackwell) with llm-compressor → compressed-tensors (nvfp4-pack-quantized); validated served on a DGX Spark (GB10, Blackwell sm_120).
For the optimal ~1.9× DFlash deploy (DGX Spark, with the speculative drafter + stability caps) follow the DGX Spark QuickStart. --served-model-name takes a list of aliases — name it after the model your clients already request for a drop-in cutover.
NVFP4 needs a Blackwell GPU (B200 / sm_100, or GB10 / sm_120). Reasoning model: every turn opens <think>…</think>. Sampling: temperature 0.6, top_p 0.95, top_k 20. Vision is intact; KV cache stays BF16 on vision deploys. On the DGX Spark's unified memory keep --gpu-memory-utilization at 0.6-0.7; above ~0.8 the shared CPU+GPU pool page-thrashes. Discrete-VRAM GPUs can run higher.
Speed (DGX Spark / GB10, measured)
Plain decode (NVFP4, no speculative decoding):
Concurrency
c=1
c=8
c=16
c=32
tok/s (aggregate)
39
221
376
539
c=1 TTFT ≈ 65 ms. Throughput scales cleanly with concurrency.
DFlash speculative decoding (~1.9× single-stream)
Because this model is a light RL post-train of Qwen3.6-35B-A3B, a Qwen3.6-35B-A3B DFlash drafter is dimensionally compatible (hidden 2048, vocab 248320; it taps target layers [1,10,19,28,37], all valid here) and accelerates this NVFP4 build out-of-the-box — pos-0 acceptance ≈ 70 %, mean acceptance length ≈ 3.5. Single-stream (c=1) decode by num_speculative_tokens:
n
4
6
8
11
c=1 tok/s
67.5
73.8 (1.89×)
69.3
66.7
Optimal n=6 → 73.8 tok/s (1.89× over plain). Add to the serve command:
The validated safe operating envelope on the GB10 — stable through c=16, peak 80 / 121 GB (≈41 GB headroom):
Concurrency
aggregate tok/s
c=1 (per user)
70 (1.78× vs plain 39)
c=8
203
c=16
456
⚠️ Stability — cap --max-num-seqs at ~16 with DFlash on the Spark. DFlash's speculative-verify buffers are not fully counted by --gpu-memory-utilization; at a cap of 32–64 they exhaust the 121 GB unified pool and hard-crash the box (kernel NVRM … NV_ERR_NO_MEMORY). For higher batch throughput, run plain NVFP4 (no DFlash) — it scales to higher concurrency safely. Use --gpu-memory-utilization 0.6 with DFlash for headroom.
Aggregate throughput vs concurrency
Prefix caching — prefill on repeated context (multi-turn / agentic)
--enable-prefix-caching (on by default in the AEON image) skips re-prefilling shared context — a major win for multi-turn chat, agentic loops, and shared system prompts. Re-prefill of a repeated context (cache miss) vs a cache hit, measured on a DGX Spark:
Context
No cache (re-prefill)
Cache hit
Speedup
1.2k
422 ms
86 ms
4.9×
4.9k
869 ms
154 ms
5.6×
9.7k
1,716 ms
103 ms
16.6×
14.6k
2,588 ms
202 ms
12.8×
Cache-hit prefill stays ~flat (~100–200 ms) while cold prefill grows with context → the longer the shared context, the bigger the win. Validated coherent: a cache-hit long-context answer is correct and identical to the full-prefill answer (the GatedDeltaNet align prefix-cache path does not corrupt the recurrent state). Composes with DFlash (DFlash speeds decode; prefix caching speeds prefill).
Prefix caching scaling
Variants
-BF16 — full precision (~66 GB), source of truth, runs on any vLLM.
FP8 is not viable on this hybrid GatedDeltaNet-MoE (W8A8 degrades coherence; W8A16 has no ScaledMM kernel) — NVFP4 is the low-precision path.
User Responsibility & Arbitration Clause
This is an uncensored model. Safety refusals have been removed, so it will generate content the base model would refuse — including instructions for harmful tools, chemicals, biological agents, or exploit code; depictions of violence, self-harm, or graphic sexuality; content that may be illegal in one or more jurisdictions; and content a reasonable person may find offensive, distressing, or morally repugnant. It makes no internal judgement about whether to comply — it complies, and your prompts are the sole determinant of what comes out. Ornith is additionally a state-of-the-art agentic-coding model: its outputs are routinely tool calls and code that downstream systems execute, so an unsafe, malicious, or merely mistaken instruction can become a real-world action with real side effects — potentially without a human in the loop. Wielding it requires a different operational stance: you, not the model, are the safety layer.
By accessing, downloading, using, running inference on, fine-tuning, merging, quantizing, distributing, integrating, or otherwise interacting with this model, you acknowledge and agree to the following:
Sole Responsibility. You, the user, are solely and exclusively responsible for (a) every prompt you or your downstream system issue to this model, (b) every response this model produces in reply, (c) every downstream action taken by you, your systems, your agents, or your users in reliance on those responses, and (d) any harm — direct, indirect, consequential, foreseeable, or otherwise — that results from any of the above.
No Warranty. This model is provided strictly "AS IS", without warranty of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, non-infringement, safety, alignment, factual accuracy, or legal compliance in any jurisdiction. No contributor, author, publisher, or hosting platform assumes liability of any kind for outputs or downstream use.
Legal Compliance. You are responsible for ensuring that your use of this model complies with all applicable laws, regulations, terms of service, industry codes of conduct, professional ethical standards, and organizational policies in every jurisdiction in which you operate or in which your outputs may be received. The unaligned nature of this model does not grant you any legal authorization you did not already have.
Operational Safety Layer. An uncensored model is not a toy. You are expected to implement appropriate downstream safety layers proportionate to your deployment context, including but not limited to: input validation, output filtering, content moderation, audit logging, rate limiting, access controls, and human-in-the-loop review for high-risk workflows — especially any agentic or autonomous-execution pipeline, where this model's tool calls and generated code run with real side effects. A production deployment of this model without such layers is unsafe by construction and is not a supported use case.
Heightened Duty of Care. The absence of internal refusal behavior means the duty of care that would ordinarily rest partly with the model rests entirely with you. You are expected to exercise greater — not lesser — caution, forethought, and ethical discipline when operating this model than you would operate a base aligned model. If you are uncertain whether your contemplated use is ethical, legal, or wise, the correct action is to not make the request.
No Endorsement of Outputs. The authors, contributors, and publishers of this model do not endorse, adopt, or take responsibility for any specific output this model produces. Outputs are a stochastic function of the prompt, the weights, and the sampler state — not a statement of position by any human.
Arbitration. Any dispute, claim, or controversy arising out of or relating to the use of this model, its outputs, or this clause shall be resolved through binding individual arbitration under the rules of a mutually agreed arbitration body (or, absent agreement, the American Arbitration Association's Consumer Arbitration Rules), waiving any right to a jury trial, class action, representative action, or consolidated proceeding. Venue shall be the jurisdiction of the disputing party bringing the claim. Costs and attorneys' fees shall be allocated per the applicable arbitration rules. This clause does not expand, and where legally prohibited does not establish, any liability in the other direction; it limits how the user may proceed when alleging harm tied to their own use of this model.
Indemnification. You agree to indemnify, defend, and hold harmless the authors, contributors, and publishers of this model from and against any claims, damages, losses, liabilities, costs, and expenses (including reasonable attorneys' fees) arising from or related to your use of the model or your breach of this clause.
Severability. If any provision of this clause is held unenforceable in a given jurisdiction, the remaining provisions remain in full force in that jurisdiction, and the unenforceable provision is replaced by the closest enforceable equivalent consistent with the original intent.
Acceptance. Your use of this model constitutes your acceptance of this clause in full. If you do not accept, do not use the model.
This model is a tool with no opinions of its own. You supply the opinions. You supply the judgement. You supply the ethics. The outputs carry your fingerprints, not the model's.
