What this quant is (and isn't)
A surgically-scoped MLP-only, weight-only NVFP4 quant:
Table with columns: Component, Precision, Why| 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)
Table with columns: Metric, BF16 parent, This NVFP4| Metric | BF16 parent | This NVFP4 |
|---|
| Agentic/coding pass@1 (18-task self-contained probe) | 0.833 (15/18) | 0.833 (15/18) — identical |
| Refusals (diverse harmful prompts) | 0 | 0 — abliteration fully survived FP4 |
| Coherence (benign + harmful, long gen) | clean | 0/10 degenerate |
| Loads with 0 skipped weights | — | ✅ |
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).
Quickstart (vLLM, Blackwell required)
vllm serve AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4 \
--served-model-name ornith --quantization compressed-tensors \
--max-model-len 262144 --gpu-memory-utilization 0.85 \
--max-num-batched-tokens 16384 --mamba-cache-dtype float32 \
--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
--limit-mm-per-prompt '{"image":4,"video":2}' --mm-encoder-tp-mode data \
--attention-backend flash_attn \
--enable-chunked-prefill --enable-prefix-caching --trust-remote-code
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.
Speed (DGX Spark / GB10, measured)
Plain decode (NVFP4, no speculative decoding):
Table with columns: Concurrency, c=1, c=8, c=16, c=32| 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:
Table with columns: n, 4, 6, 8, 11| 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:
--speculative-config '{"method":"dflash","model":"/path/to/qwen3.6-35b-a3b-dflash-drafter","num_speculative_tokens":6}'
DFlash requires BF16 KV (already forced by the vision tower). The speedup is largest at c=1 and tapers under high concurrency as the GPU saturates.
Two wins, measured separately (DGX Spark, single-stream)
Same Spark, same prompts. BF16 → NVFP4 isolates the quantization win; NVFP4 → NVFP4 + DFlash (on the AEON container) isolates the optimization win:
Table with columns: Workload, BF16 · stock vLLM, NVFP4 · stock vLLM, NVFP4 + DFlash · AEON| Workload | BF16 · stock vLLM | NVFP4 · stock vLLM | NVFP4 + DFlash · AEON |
|---|
| Coding | 30.8 tok/s · 237 ms | 38.5 · 70 ms | 77.1 · 94 ms |
| Reasoning | 30.6 · 247 ms | 38.4 · 77 ms | 107.0 · 93 ms |
| Math | 30.5 · 221 ms | 38.3 · 72 ms | 119.0 · 88 ms |
| Prose | 30.4 · 193 ms | 38.3 · 69 ms |
- Quantization (BF16 → NVFP4, same stock vLLM): ~1.25× decode, 3.2× faster TTFT (≈230 → 72 ms), 1.5× prefill — smaller weights, lighter prefill.
- Optimization (NVFP4 → + DFlash, AEON container): ~2.4× decode — DFlash speculative decoding, biggest on predictable domains (math/reasoning).
- Combined vs a naive BF16 deploy: 3.05× decode · 2.5× TTFT · 2.75× prefill.
See the DGX Spark QuickStart.
Concurrency / throughput (DGX Spark, NVFP4 + DFlash n=6, max-num-seqs 16)
The validated safe operating envelope on the GB10 — stable through c=16, peak 80 / 121 GB (≈41 GB headroom):
Table with columns: Concurrency, aggregate tok/s| 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.
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:
Table with columns: Context, No cache (re-prefill), Cache hit, Speedup| 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).
Variants
- -BF16 — full precision (~66 GB), source of truth, runs on any vLLM.
- -NVFP4 (this repo) — 4-bit, ~23.7 GB, near-lossless, Blackwell only.
- 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.
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.
☕ Support the work
If this release is useful, tips fuel more compute and more open models — thank you. QR codes on the profile »
- ₿ BTC —
bc1q09xmzn00q4z3c5raene0f3pzn9d9pvawfm0py4
- Ξ ETH —
0x1512667F6D61454ad531d2E45C0a5d1fd82D0500
- ◎ SOL —
DgQsjHdAnT5PNLQTNpJdpLS3tYGpVcsHQCkpoiAKsw8t
- ⓜ XMR —
836XrSKw4R76vNi3QPJ5Fa9ugcyvE2cWmKSPv3AhpTNNKvqP8v5ba9JRL4Vh7UnFNjDz3E2GXZDVVenu3rkZaNdUFhjAvgd
Provenance & credits
License: MIT (inherited from the base model).