Ornith-1.0-35B B70 Turbo GGUF
Ornith-1.0-35B B70 Turbo artworkThis repository publishes a
Q5_K_M GGUF for
deepreinforce-ai/Ornith-1.0-35B
plus the tested Intel Arc Pro B70 serving recipe.
The "Turbo" part is the serving stack, not a behavior-changing fine-tune:
- same base model semantics as Ornith-1.0-35B
- Q5_K_M GGUF quantization
- llama.cpp SYCL runtime tuned for Intel Arc Pro B70
- no safeguard edits, refusal edits, or new training
- no speculative-head artifact in this release
For the benchmark package, raw tables, charts, and example games, see:
newjordan/Ornith-1.0-35B-B70-Turbo.
Files
| File | Purpose |
|---|
ornith-1.0-35b-Q5_K_M.gguf | Q5_K_M GGUF model file |
ornith-1.0-35b-Q5_K_M.gguf.sha256 | SHA-256 checksum |
Recommended B70 Serve Config
Agent-fleet default:
1GGML_SYCL_DISABLE_DNN=1 ONEAPI_DEVICE_SELECTOR=level_zero:gpu \
2llama-server \
3 -m ornith-1.0-35b-Q5_K_M.gguf \
4 --alias ornith-1.0-35b-turbo \
5 -ngl 99 -fa on -ctk f16 -ctv f16 \
6 -c 131072 -np 32 -b 8192 -ub 4096 \
7 --host 0.0.0.0 --port 8092 --jinja
Single deep-agent profile:
1GGML_SYCL_DISABLE_DNN=1 ONEAPI_DEVICE_SELECTOR=level_zero:gpu \
2llama-server \
3 -m ornith-1.0-35b-Q5_K_M.gguf \
4 --alias ornith-1.0-35b-turbo \
5 -ngl 99 -fa on -ctk f16 -ctv f16 \
6 -c 262144 -np 1 -b 8192 -ub 4096 \
7 --host 0.0.0.0 --port 8092 --jinja
Avoid -np >= 56 on a single B70 in the measured fleet harness; it thrashed or
timed out under the tested workload.
Measured Performance
Hardware and stack:
- GPU: Intel Arc Pro B70, 30.3 GiB, 230 W
- Runtime: llama.cpp SYCL
- Quant: Q5_K_M
- KV cache in the final ship profile: f16
The benchmark split compares:
upstream: mainline llama.cpp, default flagsup+flags: mainline llama.cpp plus B70 runtime flagsTurbo: B70 fusion build plus the same runtime flags
Prefill
The tuned runtime configuration provides most of the prefill win.
| Prompt tokens | upstream | up+flags | Turbo | Total win |
|---|
| 805 | 1075 | 1378 | 1386 | 1.29x |
| 3313 | 1074 | 1840 | 1846 | 1.72x |
| 6963 | 1031 | 1741 | 1734 | 1.68x |
| 14563 | 959 | 1538 | 1531 | 1.60x |
| 29713 | 825 | 1208 | 1205 | 1.46x |
| 61341 | 628 | 828 | 826 | 1.32x |
| 129325 | 413 | 489 | 488 | 1.18x |
Single-Stream Decode
The B70 fusion build provides the single-stream decode gain.
| Context depth | upstream | up+flags | Turbo | Total win |
|---|
| 805 | 81.7 | 81.8 | 93.5 | 1.14x |
| 3313 | 80.0 | 79.8 | 91.2 | 1.14x |
| 6963 | 77.5 | 77.4 | 88.2 | 1.14x |
| 14563 | 72.9 | 72.7 | 82.3 | 1.13x |
| 29713 | 66.5 | 66.1 | 74.1 | 1.11x |
| 61341 | 55.7 | 55.5 | 61.1 | 1.10x |
| 129325 | 41.4 | 41.3 | 44.2 | 1.07x |
Fleet Decode
Aggregate decode with a synthetic 2048+256 workload:
| Agents | upstream | up+flags | Turbo | Total win |
|---|
| 1 | 78.5 | 78.9 | 86.2 | 1.10x |
| 4 | 85.7 | 118.8 | 120.8 | 1.41x |
| 8 | 91.4 | 132.8 | 132.8 | 1.45x |
| 16 | 98.1 | 132.3 | 132.5 | 1.35x |
| 24 | 103.2 | 143.8 | 139.0 | 1.35x |
| 32 | 112.0 | 149.1 | 149.1 | 1.33x |
| 48 | 122.9 | 160.8 | 157.0 | 1.28x |
| 56 | 126.6 | 161.8 | 160.2 | 1.27x |
Accuracy Snapshot
These are reference lm-eval results carried over from the local benchmark set;
serving changes are lossless for the same GGUF.
| Benchmark | Score |
|---|
| GSM8K | 97.0 |
| HellaSwag | 82.1 |
| Winogrande | 71.6 |
| ARC-Challenge | 49.2 |
| MMLU | 41.1 |
| TruthfulQA-MC1 | 35.7 |
| Wikitext2 PPL | 6.36 |
Notes
- This release is for llama.cpp-compatible GGUF runtimes.
- The best measured B70 profile uses f16 KV. q8_0 KV was slower at long depth
on this SYCL backend in local testing.
- The route-aware DeepSpec/Eagle3 speculative-head work is tracked separately
and is not included in this artifact.
