Ornith-1.0-35B — AEON Ultimate Uncensored

Uncensored / abliterated build of deepreinforce-ai/Ornith-1.0-35B, DeepReinforce's SOTA agentic-coding MoE — refusals removed, capability preserved.

BF16 (~66 GB, any vLLM)	AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16
NVFP4 (~23.7 GB, near-lossless, Blackwell)	AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4
License	MIT (inherited from base)

TL;DR

• 0 / 80 refusals (0.0%) on diverse harmful prompts (base ≈ 94/100) — fully uncensored.
• 0 coding-capability loss — agentic pass@1 0.833, identical to the base family-by-family. Near-lossless (KL ≈ 0.0014).
• Hybrid qwen3_5_moe: 40 layers (30 GatedDeltaNet + 10 full-attn), 256 experts + shared (A3B), vision, 256K ctx, thinking model.
• NVFP4 + DFlash on a DGX Spark: 3.05× faster decode (avg 93 vs 30 tok/s) and 2.75× prefill vs a stock-vLLM BF16 deploy.

Performance — DGX Spark (GB10), single-stream

Three stacks on the same Spark, so the quantization win and the optimization win are visible separately:

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	70.3 · 91 ms
Avg decode	30.6	38.4	93.3
Prefill	3,517 tok/s	5,203	9,661

• Quantization (BF16→NVFP4): ~1.25× decode, 3.2× faster TTFT, 1.5× prefill, 64% smaller.
• Optimization (NVFP4→+DFlash, AEON container): ~2.4× decode.
• Combined: 3.05× decode · 2.5× TTFT · 2.75× prefill vs a naive BF16 deploy.

Concurrency (NVFP4 + DFlash n=6, max-num-seqs 16 — the safe envelope): c=1 70 tok/s (1.78× vs plain 39), c=8 203, c=16 456 aggregate; stable, peak 80/121 GB.

⚠️ DFlash + Spark unified memory: keep --max-num-seqs ≤ 16 with DFlash. Its speculative-verify buffers aren't 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) instead of raising the cap.

▶ Benchmark animation (MP4) · 📖 DGX Spark QuickStart (optimal settings: NVFP4 + DFlash n=6, --gpu-memory-utilization 0.6, vision + tool-calling flags)

Quickstart (vLLM)

vllm serve AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16 \
  --served-model-name ornith --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

Thinking model (<think>…</think> every turn). Sampling: temperature 0.6, top_p 0.95, top_k 20. Vision intact (BF16 KV on vision deploys). See serve_ornith.sh.

How it was built (4 stages)

1. SSM conv1d outlier repair — rescale outlier blocks (layers 36/37) pre-abliteration to prevent coherence collapse.
2. Abliteration — abliterix v1.9: grimjim norm-preserving biprojection + Expert-Granular Abliteration across all 256 fused experts + shared expert + router suppression; Optuna refusals-vs-KL search. Q/K/V untouched (attn_output_gate); GatedDeltaNet/SSM internals + vision tower not modified. Recipe: config/ornith_35b.toml.
3. Gentle-knee selection — the lowest-refusal trial was over-abliterated (word-salad on real generation). Shipped a 170× lighter expert edit for the same refusal removal, verified coherent.
4. Export + quantize. Export bf16 on a box where 70 GB fits in memory (GB10's 121 GB unified pool cannot edit-in-memory). NVFP4 (shipped): MLP-only weight-only NVFP4 (W4A16) via llm-compressor → compressed-tensors — experts + shared-MLP in NVFP4, attention + GatedDeltaNet + vision + gates + embeds in BF16. Quantized on a B200 (Blackwell), validated served on a DGX Spark (GB10): identical to BF16 — agentic-coding 0.833 (15/18), 0 refusals, 0 degenerate, ~23.7 GB. Weight-only (BF16 activations) is the key: the NVFP4 reasoning penalty comes from FP4 activations (W4A4), which W4A16 avoids. FP8 is not viable on this hybrid arch (W8A8 degrades coherence; W8A16 has no ScaledMM kernel) — NVFP4 is the low-precision path.

Validation

Metric	Base Ornith-1.0-35B	This model
Refusals (80 harmful: CBRN/cyber/weapons/self-harm)	high	0 / 80 (0.0%)
Agentic/coding pass@1 (18-task probe)	0.833	0.833 (identical)
First-token KL vs base	—	~0.0014

Base capability context (preserved, per the coding-delta check): Terminal-Bench 2.1 = 64.2, SWE-bench Verified = 75.6.

User Responsibility & Arbitration Clause

This is an uncensored, agentic model — safety refusals are removed (it will comply with harmful requests), and its outputs are routinely tool calls and code that downstream systems execute. You, not the model, are the safety layer. By using it you accept: sole responsibility for all prompts, outputs, and downstream actions and any resulting harm; that it is provided "AS IS" with no warranty of any kind; that you must comply with all applicable laws in every relevant jurisdiction; that you will implement your own downstream safety layers (input/output filtering, content moderation, audit logging, access controls, and human-in-the-loop review — especially for agentic/autonomous-execution pipelines); no endorsement of any output by the authors; and binding individual arbitration, indemnification, and severability terms.

Read the full, governing User Responsibility & Arbitration Clause on the BF16 model card before downloading or deploying.

☕ 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

Base deepreinforce-ai/Ornith-1.0-35B · driver abliterix / heretic · methods grimjim (NPBA), Arditi et al. 2024, FernflowerAI (SSM repair) · build AEON-7.