Ornith-1.0-9B-DSpark

Ornith-1.0-9B-DSpark is a production-qualified DSpark speculative-decoding draft for deepreinforce-ai/Ornith-1.0-9B.

This repository contains the draft/speculator only. It is not a standalone language model. The draft must be paired with the exact Ornith verifier revision described in candidate-manifest.json. When served through speculative decoding, the verifier remains the source of truth for outputs and model capability; this draft proposes token blocks that the verifier accepts or rejects.

What Is Included

File	Purpose
model.safetensors	BF16 DSpark draft weights
config.json	Draft architecture, target-layer mapping, block size, and calibrated confidence temperatures
candidate-manifest.json	Draft and verifier checksums plus repeated acceptance runs
qualification.json	Final 17-gate qualification result
serving/	vLLM configuration, router, and compatibility patches required for the qualified setup
LICENSE	MIT license

Large training logs, intermediate checkpoints, handoff notes, local deployment experiments, and raw evaluation dumps are intentionally not included.

Results

All measurements were run against the matched Ornith verifier on NVIDIA B200 hardware with the serving configuration represented in this repository.

Metric	Result
Qualification gates	17 / 17 passed
Mean accepted length, 3 frozen-candidate repeats	4.5865
Accepted length range	4.5619 - 4.6053
DSpark paper reference used for comparison	4.8133
Paper-reference ratio	95.3%
GSM8K target-only correct	925 / 1,024
GSM8K DSpark correct	926 / 1,024
One-sided 95% accuracy lower bound	-0.64 percentage points
Online speedup, concurrency 1	2.46x
Online speedup, concurrency 8	1.65x
Online speedup, concurrency 32	1.22x
Online draft acceptance, C1 / C8 / C32	41.4% / 39.2% / 38.4%
Sampling-distribution excess TV	0.0
Non-low-margin top-token mismatches	0
Stability soak	1,024 / 1,024 successful
Long-context validation	8K, 32K, 64K, 128K, 256K passed

The speed figures are specific to the tested Ornith/B200/vLLM stack. Different GPUs, batching policies, vLLM versions, or target revisions require fresh profiling and qualification.

Architecture

Property	Value
Draft architecture	Qwen35DSparkModel
Target verifier	deepreinforce-ai/Ornith-1.0-9B
Draft layers	5
Target feature layers	1, 8, 15, 22, 29
Hidden size	4,096
Block size	7
Markov head	Vanilla, rank 256
Confidence head	Enabled, Markov-conditioned
Calibration	Sequential temperature scaling
Vocabulary	248,320
Weight dtype	BF16
Context policy	DSpark up to 4,096 input-plus-output tokens; target-only fallback above that

Download

hf download stanleyphoong/Ornith-1.0-9B-DSpark \
  --local-dir ./Ornith-1.0-9B-DSpark

hf download deepreinforce-ai/Ornith-1.0-9B \
  --local-dir ./Ornith-1.0-9B

The draft repository is about 6.2 GiB. The Ornith verifier is a separate download and is not duplicated here.

Serving Notes

The qualified runtime uses DSpark support that is not assumed to exist in an unmodified stable vLLM release. The files under serving/ provide the tested server specifications, router, DSpark config builder, and vLLM compatibility patches.

Use the DSpark route only for requests whose prompt plus requested output is at most 4,096 tokens. Larger requests should be routed to target-only Ornith, which preserves the validated 262,144-token context path and avoids measured long-context slowdown in the draft route.

Required compatibility rules:

1. Use the exact target weights/checksums in candidate-manifest.json.
2. Keep the calibrated confidence_temperatures in config.json.
3. Install or mount every file in serving/vllm-patches/ at the paths expected by the server specification.
4. Re-profile before changing GPU type, vLLM version, batching, maximum concurrency, or verifier revision.
5. Re-run qualification after any runtime or target-model change.

Losslessness

Here, "lossless" follows the speculative-decoding definition: verifier-side accept/reject sampling preserves the target model distribution. Qualification used first-token and teacher-forced distribution checks, empirical temperature-1 sampling comparisons, paired labeled accuracy, online acceptance, load testing, memory checks, and long-context routing validation.

Bit-identical greedy strings across different execution shapes are not treated as the proof of losslessness because near-tied logits can select different tokens under different batching shapes. The hard gates are recorded in qualification.json.

Limitations

• This draft is verifier-specific and should not be used as a universal speculator.
• The published artifact is a draft model, not a chat model or standalone generator.
• Text generation was qualified; multimodal speculative decoding was not.
• The DSpark route is validated for the 4K routing budget described above.
• New hardware, runtime patches, target revisions, or routing policies require requalification.

Citation

@misc{ornith_9b_2026,
  title  = {Ornith-1.0-9B: Agentic Coding, Open to All},
  author = {{DeepReinforce Team}},
  year   = {2026},
  url    = {https://deep-reinforce.com/ornith_1_0.html}
}

Please also cite the DSpark/DeepSpec work when using this speculator.