Ornith-1.5-122B-EXL3-SAGE-2.58bpw API & Inference Endpoint

Pack Summary

Table with columns: Field, Value
Field	Value
Base Model	`ornith-ai/Ornith-1.5-122B` (`125.09 B` parameters, BF16)
Architecture	`Qwen3_5MoeForConditionalGeneration` (48 layers: 36 Gated DeltaNet + 12 full attention, 256 routed experts top-8 + 1 shared expert, 1 native MTP layer, 27-layer vision tower)
Format	EXL3 (`quant_method: exl3`, codebook `mul1`)
Average BPW	2.58 bpw
Pack Size	41.00 GiB language model weights (`39.58 GiB` GPU-resident LM + `1.42 GiB` `embed_tokens`) / 42.01 GiB total on disk across 6 shards (`model-00001-of-00006.safetensors` .. `model-00006-of-00006.safetensors`)
Context Window	`262,144` tokens (`max_position_embeddings`, extendable to `1M` via YaRN)
Target Hardware	Single 48 GB GPU or 128 GB unified-memory system (NVIDIA DGX Spark `GB10`)
License	MIT

Measured Speeds on 1× NVIDIA DGX Spark (`GB10`, `sm_121`, 128 GB Unified)

Measured end-to-end with vLLM 0.32.0 + vcruz305/vllm-exl3 (using vcruz305/Ornith-1.5-122B-EXL3-DGX-Spark-recipe) and REXL3:

Table with columns: Engine / Workload (GB10, 1× DGX Spark), Decode (tok/s), Per-Stream (tok/s), Prefill (tok/s), TTFT (ms), Notes
Engine / Workload (`GB10`, 1× DGX Spark)	Decode (`tok/s`)	Per-Stream (`tok/s`)	Prefill (`tok/s`)	TTFT (`ms`)	Notes
`vLLM 0.32.0 + vllm-exl3` (`C = 1`, MTP `k=1` on Code)	32.60 (32.99 peak)	32.60	—	124.2	`88.6%` draft acceptance on code; `30.38 tok/s` mean across all suites
`vLLM 0.32.0 + vllm-exl3` (`C = 1`, No Draft)	30.51 (30.57 peak)	30.51	1,119.1 (2K)	109.1	`VLLM_EXL3_COOP=1` + `VLLM_EXL3_COALESCE_SHARDS=1` (`2.0 s` CUDA graph capture)
`vLLM 0.32.0 + vllm-exl3` (`C = 4` Concurrent)	63.07	17.22	—	238.1	2.07× aggregate throughput
`vLLM 0.32.0 + vllm-exl3` (`C = 8` Concurrent)	89.21	12.02	—	428.7	2.92× aggregate throughput
`vLLM 0.32.0 + vllm-exl3` Prefill `2,048` (Cold / Cached)	—	—	1,119.1 / 1,408.7	1,830.1 / 1,453.9	1.26× prefix-cache speedup
`vLLM 0.32.0 + vllm-exl3` Prefill `4,096` (Cold / Cached)	—	—	1,033.8 / 2,210.9	3,962.1 / 1,852.6	2.14× prefix-cache speedup
`vLLM 0.32.0 + vllm-exl3` Prefill `8,192` (Cold / Cached)	—	—	991.6 / 4,449.5	8,261.2 / 1,841.1	4.49× prefix-cache speedup
Reference: `REXL3` Native Engine (`C = 1` / `C = 8`)	30.58 / 90.82	30.58 / 11.35	—	91.2	`9.12 s` cold boot (`41.00 GiB` LM weights)

9-Suite (`353`/`356`-Item) Downstream Benchmark vs. `4.0 bpw`

Measured on the same DGX Spark (GB10) under identical deterministic prompts and evaluation harnesses:

Table with columns: Benchmark Suite, Items, This Pack (vcruz305 SAGE 2.58 bpw, 41.00 GiB LM), 4.0 bpw EXL3 (59.00 GiB LM), Delta (SAGE 2.58 vs 4.0 bpw)
Benchmark Suite	Items	This Pack (`vcruz305` SAGE 2.58 bpw, `41.00 GiB` LM)	`4.0 bpw` EXL3 (`59.00 GiB` LM)	Delta (`SAGE 2.58` vs `4.0 bpw`)
MMLU-Pro (STEM)	100	94 / 100 (`94.00%`)	96 / 100 (`96.00%`)	-2 (`-2.00%`)
LiveCodeBench (Matched / Full)	17 / 20	13 / 17 (`76.47%`) / 15 / 20 (`75.00%`)	12 / 17 (`70.59%`) / 14 / 20 (`70.00%`)	+1 (`+5.88%` / `+5.00%`)
MBPP+	60	53 / 60 (`88.33%`)	51 / 60 (`85.00%`)	+2 (`+3.33%`)
HumanEval+	40	39 / 40 (`97.50%`)	39 / 40 (`97.50%`)	0 (`0.00%`)
SciCode	10	5 / 10 (`50.00%`)	4 / 10 (`40.00%`)	+1 (`+10.00%`)
MATH-500	40	38 / 40 (`95.00%`)	39 / 40 (`97.50%`)	-1 (`-2.50%`)
MGSM (8 Languages)	40	40 / 40 (`100.00%`)	40 / 40 (`100.00%`)	0 (`0.00%`)
BFCL (Function Calling)	40	37 / 40 (`92.50%`)	37 / 40 (`92.50%`)	0 (`0.00%`)
Multi-step Tool Chains	6	5 / 6 (`83.33%`)	5 / 6 (`83.33%`)	0 (`0.00%`)
Coding Subtotal (LCB+MBPP+HE+SciCode)	130	112 / 130 (`86.15%`)	108 / 130 (`83.08%`)	+4 (`+3.08%`)
OVERALL TOTAL (353-Item Matched)	353	324 / 353 (`91.78%`)	323 / 353 (`91.50%`)	+1 (`+0.28%`)
OVERALL TOTAL (356-Item Full)	356	326 / 356 (`91.57%`)	325 / 356 (`91.29%`)	+1 (`+0.28%`)
Loaded LM Weight Footprint	—	41.00 GiB (fits 1× 48 GB GPU)	59.00 GiB (requires 80 GB+)	-18.00 GiB (`-30.5%` smaller)
Cold Boot Time (`REXL3`, `GB10`)	—	9.12 s	19.23 s	2.11× faster
`C = 8` Aggregate Decode (`GB10`)	—	89.21 tok/s (`vLLM`) / 90.82 tok/s (`REXL3`)	83.11 tok/s	+9.3% faster

Fidelity vs. Unquantized `ornith-ai/Ornith-1.5-122B` BF16 Master

Measured directly against the unquantized ornith-ai/Ornith-1.5-122B BF16 master across held-out evaluation splits over the full 248,320-token vocabulary:

Table with columns: Evaluation Split / Subset, Positions, Top-1 Agreement, Top-5 Agreement, Mean KLD (nats), Median KLD (nats), Mean TV, BF16 Master PPL, SAGE 2.58bpw PPL, ΔPPL (%)
Evaluation Split / Subset	Positions	Top-1 Agreement	Top-5 Agreement	Mean KLD (nats)	Median KLD (nats)	Mean TV	BF16 Master PPL	SAGE 2.58bpw PPL	$Δ$ PPL (%)
`e2e_eval` (Answer-Only)	5,256	92.10%	99.24%	0.10845	0.00175	0.07938	2.2139	2.2984	+3.82%
`e2e_eval` (Non-Header)	16,120	91.12%	99.09%	0.16414	0.00383	0.09176	2.6812	2.7426	+2.29%
`e2e_eval` — Code (HumanEval)	4,038	94.43%	99.21%	0.14847	0.00041	0.06060	1.6854	1.7555	+4.16%
`e2e_eval` — Reasoning (MATH-500)	4,036	91.18%	99.28%	0.15074	0.00516	0.08967	3.0384	2.9640	-2.45%
`e2e_eval` — Math (GSM8K)	4,018	90.24%	99.08%	0.20349	0.00491	0.10177	2.3485	2.4310	+3.51%
`e2e_eval` — General (WikiText-2)	4,028	88.60%	98.81%	0.15403	0.02442	0.11510	4.2996	4.4758	+4.10%
`heldout_8x1024` (Answer-Only)	2,664	90.58%	99.29%	0.09522	0.00299	0.08531	2.4456	2.5333	+3.59%
`heldout_8x1024` — Code	2,026	94.82%	99.41%	0.14258	0.00035	0.05740	1.6215	1.6995	+4.81%
`heldout_8x1024` — Reasoning	2,024	91.90%	99.11%	0.13906	0.00182	0.08081	2.4344	2.4775	+1.77%
`indep_8x1024` (Non-Header)	8,056	90.83%	99.14%	0.14847	0.00575	0.09303	2.8513	2.8938	+1.49%
`indep_8x1024` — Code	2,012	94.43%	99.25%	0.14510	0.00039	0.05797	1.7349	1.7913	+3.25%
`indep_8x1024` — Reasoning	2,026	93.09%	99.51%	0.08431	0.00607	0.07276	2.7303	2.7202	-0.37%
`indep_8x1024` — Math	2,010	90.00%	99.05%	0.17364	0.00998	0.10259	2.9910	3.0177	+0.89%

Quickstart: Serving with `vLLM 0.32.0` + `vcruz305/vllm-exl3`

Full scripts, preflight checks, and benchmarks are available in the companion recipe repository:
👉 https://github.com/vcruz305/Ornith-1.5-122B-EXL3-DGX-Spark-recipe

bash

git clone https://github.com/vcruz305/Ornith-1.5-122B-EXL3-DGX-Spark-recipe.gitcd Ornith-1.5-122B-EXL3-DGX-Spark-recipe # 1. Create the vLLM environment (vllm==0.32.0, vcruz305/vllm-exl3 with -DP2B_CB=2, vision patch).bash vllm-plugin/setup_env.sh # 2. Download this pack.HF_XET_HIGH_PERFORMANCE=1 ~/venvs/vllm-exl3/bin/hf download \  vcruz305/Ornith-1.5-122B-EXL3-SAGE-2.58bpw \  --local-dir ~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw # 3. Prepare a non-destructive symlinked view at ~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw-vllm.PACK_DIR=~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw \  PLUGIN_REPO=/path/to/vllm-exl3 \  bash vllm-plugin/prepare_pack.sh # 4. Launch vLLM (--quantization exl3, 262,144 context, VLLM_EXL3_COOP=1, VLLM_EXL3_COALESCE_SHARDS=1).source ~/venvs/vllm-exl3/bin/activatebash vllm-plugin/serve.sh

Enable MTP speculative decoding (32.60 tok/s / 32.99 peak on code):

bash

SPEC_CONFIG='{"method":"mtp","num_speculative_tokens":1}' bash vllm-plugin/serve.sh

Enable the 27-layer multimodal vision tower:

bash

LANGUAGE_MODEL_ONLY=0 bash vllm-plugin/serve.sh

About the Base Model: `ornith-ai/Ornith-1.5-122B`

(Copied from the upstream ornith-ai/Ornith-1.5-122B model card for reference)

Chirp Chirp! 🐦 We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement.

Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For more details on the task, harness, and rollout reward design, please refer to our blog.

Upstream Benchmarks (`ornith-ai/Ornith-1.5-122B`)

Recommended Sampling & Reasoning Parameters

Using Ornith-1.5-122B via the Chat Completions API

Once your vLLM server is running (http://localhost:8899/v1 via vcruz305/Ornith-1.5-122B-EXL3-DGX-Spark-recipe), talk to it with any OpenAI-compatible client:

python

from openai import OpenAI client = OpenAI(    base_url="http://localhost:8899/v1",    api_key="EMPTY",) response = client.chat.completions.create(    model="Ornith-1.5-122B-EXL3",    messages=[        {"role": "user", "content": "Write a one-line Python lambda that squares a number."}    ],    temperature=0.6,    top_p=0.95,    max_tokens=1024,) message = response.choices[0].messageprint("reasoning:", getattr(message, "reasoning_content", None))print("answer:", message.content)

Tool calling works out of the box:

python

tools = [    {        "type": "function",        "function": {            "name": "get_weather",            "description": "Get the current weather for a city",            "parameters": {                "type": "object",                "properties": {"city": {"type": "string"}},                "required": ["city"],            },        },    }] response = client.chat.completions.create(    model="Ornith-1.5-122B-EXL3",    messages=[{"role": "user", "content": "What is the weather in Paris right now?"}],    tools=tools,    tool_choice="auto",    temperature=0.6,    max_tokens=2048,) tool_call = response.choices[0].message.tool_calls[0]print(tool_call.function.name, tool_call.function.arguments)# -> get_weather {"city": "Paris"}

Agentic & Coding CLI Usage

Hermes Agent

bash

export OPENAI_BASE_URL="http://localhost:8899/v1"export OPENAI_API_KEY="EMPTY"export MODEL="Ornith-1.5-122B-EXL3"

OpenClaw

bash

export OPENAI_BASE_URL="http://localhost:8899/v1"export OPENAI_API_KEY="EMPTY"export OPENAI_MODEL="Ornith-1.5-122B-EXL3"

OpenCode

bash

# Register your local Ornith endpoint as a provider in ~/.config/opencode/opencode.json:# {#   "$schema": "https://opencode.ai/config.json",#   "provider": {#     "ornith": {#       "npm": "@ai-sdk/openai-compatible",#       "name": "Ornith (local)",#       "options": { "baseURL": "http://localhost:8899/v1", "apiKey": "EMPTY" },#       "models": { "Ornith-1.5-122B-EXL3": { "name": "Ornith-1.5-122B-EXL3" } }#     }#   }# }opencode

Citation

bibtex

@misc{ornith_1_5,    title = {{Ornith-1.5}: From Self-Scaffolding to Self-Improvement},    url = {https://ornith.ai/ornith_1_5.html},    author = {{Ornith Team}},    year = {2026}}

Field

Value

Base Model

ornith-ai/Ornith-1.5-122B (125.09 B parameters, BF16)

Architecture

Qwen3_5MoeForConditionalGeneration (48 layers: 36 Gated DeltaNet + 12 full attention, 256 routed experts top-8 + 1 shared expert, 1 native MTP layer, 27-layer vision tower)

Format

EXL3 (quant_method: exl3, codebook mul1)

Average BPW

2.58 bpw

Pack Size

41.00 GiB language model weights (39.58 GiB GPU-resident LM + 1.42 GiB embed_tokens) / 42.01 GiB total on disk across 6 shards (model-00001-of-00006.safetensors .. model-00006-of-00006.safetensors)

Context Window

262,144 tokens (max_position_embeddings, extendable to 1M via YaRN)

Target Hardware

Single 48 GB GPU or 128 GB unified-memory system (NVIDIA DGX Spark GB10)

License

MIT

Engine / Workload (GB10, 1× DGX Spark)

Decode (tok/s)

Per-Stream (tok/s)

Prefill (tok/s)

TTFT (ms)

Notes

vLLM 0.32.0 + vllm-exl3 (C = 1, MTP k=1 on Code)

32.60 (32.99 peak)

32.60

—

124.2

88.6% draft acceptance on code; 30.38 tok/s mean across all suites

vLLM 0.32.0 + vllm-exl3 (C = 1, No Draft)

30.51 (30.57 peak)

30.51

1,119.1 (2K)

109.1

VLLM_EXL3_COOP=1 + VLLM_EXL3_COALESCE_SHARDS=1 (2.0 s CUDA graph capture)

vLLM 0.32.0 + vllm-exl3 (C = 4 Concurrent)

63.07

17.22

—

238.1

2.07× aggregate throughput

vLLM 0.32.0 + vllm-exl3 (C = 8 Concurrent)

89.21

12.02

—

428.7

2.92× aggregate throughput

vLLM 0.32.0 + vllm-exl3 Prefill 2,048 (Cold / Cached)

—

1,119.1 / 1,408.7

1,830.1 / 1,453.9

1.26× prefix-cache speedup

vLLM 0.32.0 + vllm-exl3 Prefill 4,096 (Cold / Cached)

—

1,033.8 / 2,210.9

3,962.1 / 1,852.6

2.14× prefix-cache speedup

vLLM 0.32.0 + vllm-exl3 Prefill 8,192 (Cold / Cached)

—

991.6 / 4,449.5

8,261.2 / 1,841.1

4.49× prefix-cache speedup

Reference: REXL3 Native Engine (C = 1 / C = 8)

30.58 / 90.82

30.58 / 11.35

—

91.2

9.12 s cold boot (41.00 GiB LM weights)

Benchmark Suite

Items

This Pack (vcruz305 SAGE 2.58 bpw, 41.00 GiB LM)

4.0 bpw EXL3 (59.00 GiB LM)

Delta (SAGE 2.58 vs 4.0 bpw)

MMLU-Pro (STEM)

100

94 / 100 (94.00%)

96 / 100 (96.00%)

-2 (-2.00%)

LiveCodeBench (Matched / Full)

17 / 20

13 / 17 (76.47%) / 15 / 20 (75.00%)

12 / 17 (70.59%) / 14 / 20 (70.00%)

+1 (+5.88% / +5.00%)

MBPP+

53 / 60 (88.33%)

51 / 60 (85.00%)

+2 (+3.33%)

HumanEval+

39 / 40 (97.50%)

0 (0.00%)

SciCode

5 / 10 (50.00%)

4 / 10 (40.00%)

+1 (+10.00%)

MATH-500

38 / 40 (95.00%)

39 / 40 (97.50%)

-1 (-2.50%)

MGSM (8 Languages)

40 / 40 (100.00%)

0 (0.00%)

BFCL (Function Calling)

37 / 40 (92.50%)

0 (0.00%)

Multi-step Tool Chains

5 / 6 (83.33%)

0 (0.00%)

Coding Subtotal (LCB+MBPP+HE+SciCode)

130

112 / 130 (86.15%)

108 / 130 (83.08%)

+4 (+3.08%)

OVERALL TOTAL (353-Item Matched)

353

324 / 353 (91.78%)

323 / 353 (91.50%)

+1 (+0.28%)

OVERALL TOTAL (356-Item Full)

356

326 / 356 (91.57%)

325 / 356 (91.29%)

+1 (+0.28%)

Loaded LM Weight Footprint

—

41.00 GiB (fits 1× 48 GB GPU)

59.00 GiB (requires 80 GB+)

-18.00 GiB (-30.5% smaller)

Cold Boot Time (REXL3, GB10)

—

9.12 s

19.23 s

2.11× faster

C = 8 Aggregate Decode (GB10)

—

89.21 tok/s (vLLM) / 90.82 tok/s (REXL3)

83.11 tok/s

+9.3% faster

Evaluation Split / Subset

Positions

Top-1 Agreement

Top-5 Agreement

Mean KLD (nats)

Median KLD (nats)

Mean TV

BF16 Master PPL

SAGE 2.58bpw PPL

Δ

PPL (%)

e2e_eval (Answer-Only)

5,256

92.10%

99.24%

0.10845

0.00175

0.07938

2.2139

2.2984

+3.82%

e2e_eval (Non-Header)

16,120

91.12%

99.09%

0.16414

0.00383

0.09176

2.6812

2.7426

+2.29%

e2e_eval — Code (HumanEval)

4,038

94.43%

99.21%

0.14847

0.00041

0.06060

1.6854

1.7555

+4.16%

e2e_eval — Reasoning (MATH-500)

4,036

91.18%

99.28%

0.15074

0.00516

0.08967

3.0384

2.9640

-2.45%

e2e_eval — Math (GSM8K)

4,018

90.24%

99.08%

0.20349

0.00491

0.10177

2.3485

2.4310

+3.51%

e2e_eval — General (WikiText-2)

4,028

88.60%

98.81%

0.15403

0.02442

0.11510

4.2996

4.4758

+4.10%

heldout_8x1024 (Answer-Only)

2,664

90.58%

99.29%

0.09522

0.00299

0.08531

2.4456

2.5333

+3.59%

heldout_8x1024 — Code

2,026

94.82%

99.41%

0.14258

0.00035

0.05740

1.6215

1.6995

+4.81%

heldout_8x1024 — Reasoning

2,024

91.90%

99.11%

0.13906

0.00182

0.08081

2.4344

2.4775

+1.77%

indep_8x1024 (Non-Header)

8,056

90.83%

99.14%

0.14847

0.00575

0.09303

2.8513

2.8938

+1.49%

indep_8x1024 — Code

2,012

94.43%

99.25%

0.14510

0.00039

0.05797

1.7349

1.7913

+3.25%

indep_8x1024 — Reasoning

2,026

93.09%

99.51%

0.08431

0.00607

0.07276

2.7303

2.7202

-0.37%

indep_8x1024 — Math

2,010

90.00%

99.05%

0.17364

0.00998

0.10259

2.9910

3.0177

+0.89%

git clone https://github.com/vcruz305/Ornith-1.5-122B-EXL3-DGX-Spark-recipe.gitcd Ornith-1.5-122B-EXL3-DGX-Spark-recipe # 1. Create the vLLM environment (vllm==0.32.0, vcruz305/vllm-exl3 with -DP2B_CB=2, vision patch).bash vllm-plugin/setup_env.sh # 2. Download this pack.HF_XET_HIGH_PERFORMANCE=1 ~/venvs/vllm-exl3/bin/hf download \ vcruz305/Ornith-1.5-122B-EXL3-SAGE-2.58bpw \ --local-dir ~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw # 3. Prepare a non-destructive symlinked view at ~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw-vllm.PACK_DIR=~/models/Ornith-1.5-122B-EXL3-SAGE-2.58bpw \ PLUGIN_REPO=/path/to/vllm-exl3 \ bash vllm-plugin/prepare_pack.sh # 4. Launch vLLM (--quantization exl3, 262,144 context, VLLM_EXL3_COOP=1, VLLM_EXL3_COALESCE_SHARDS=1).source ~/venvs/vllm-exl3/bin/activatebash vllm-plugin/serve.sh

from openai import OpenAI client = OpenAI( base_url="http://localhost:8899/v1", api_key="EMPTY",) response = client.chat.completions.create( model="Ornith-1.5-122B-EXL3", messages=[ {"role": "user", "content": "Write a one-line Python lambda that squares a number."} ], temperature=0.6, top_p=0.95, max_tokens=1024,) message = response.choices[0].messageprint("reasoning:", getattr(message, "reasoning_content", None))print("answer:", message.content)

tools = [ { "type": "function", "function": { "name": "get_weather", "description": "Get the current weather for a city", "parameters": { "type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"], }, }, }] response = client.chat.completions.create( model="Ornith-1.5-122B-EXL3", messages=[{"role": "user", "content": "What is the weather in Paris right now?"}], tools=tools, tool_choice="auto", temperature=0.6, max_tokens=2048,) tool_call = response.choices[0].message.tool_calls[0]print(tool_call.function.name, tool_call.function.arguments)# -> get_weather {"city": "Paris"}

# Register your local Ornith endpoint as a provider in ~/.config/opencode/opencode.json:# {# "$schema": "https://opencode.ai/config.json",# "provider": {# "ornith": {# "npm": "@ai-sdk/openai-compatible",# "name": "Ornith (local)",# "options": { "baseURL": "http://localhost:8899/v1", "apiKey": "EMPTY" },# "models": { "Ornith-1.5-122B-EXL3": { "name": "Ornith-1.5-122B-EXL3" } }# }# }# }opencode

Ornith-1.5-122B-EXL3-SAGE-2.58bpw

README

Pack Summary

Measured Speeds on 1× NVIDIA DGX Spark (`GB10`, `sm_121`, 128 GB Unified)

9-Suite (`353`/`356`-Item) Downstream Benchmark vs. `4.0 bpw`

Fidelity vs. Unquantized `ornith-ai/Ornith-1.5-122B` BF16 Master

Quickstart: Serving with `vLLM 0.32.0` + `vcruz305/vllm-exl3`

About the Base Model: `ornith-ai/Ornith-1.5-122B`

Upstream Benchmarks (`ornith-ai/Ornith-1.5-122B`)

Recommended Sampling & Reasoning Parameters

Using Ornith-1.5-122B via the Chat Completions API

Agentic & Coding CLI Usage

Hermes Agent

OpenClaw

OpenCode

Citation

Explore FriendliAI today

README

Pack Summary

Measured Speeds on 1× NVIDIA DGX Spark (`GB10`, `sm_121`, 128 GB Unified)

9-Suite (`353`/`356`-Item) Downstream Benchmark vs. `4.0 bpw`

Fidelity vs. Unquantized `ornith-ai/Ornith-1.5-122B` BF16 Master

Quickstart: Serving with `vLLM 0.32.0` + `vcruz305/vllm-exl3`

About the Base Model: `ornith-ai/Ornith-1.5-122B`

Upstream Benchmarks (`ornith-ai/Ornith-1.5-122B`)

Recommended Sampling & Reasoning Parameters

Using Ornith-1.5-122B via the Chat Completions API

Agentic & Coding CLI Usage

Hermes Agent

OpenClaw

OpenCode

Citation

Ornith-1.5-122B-EXL3-SAGE-2.58bpw

README

Pack Summary

Measured Speeds on 1× NVIDIA DGX Spark (GB10, sm_121, 128 GB Unified)

9-Suite (353/356-Item) Downstream Benchmark vs. 4.0 bpw

Fidelity vs. Unquantized ornith-ai/Ornith-1.5-122B BF16 Master

Quickstart: Serving with vLLM 0.32.0 + vcruz305/vllm-exl3

About the Base Model: ornith-ai/Ornith-1.5-122B

Upstream Benchmarks (ornith-ai/Ornith-1.5-122B)

Recommended Sampling & Reasoning Parameters

Using Ornith-1.5-122B via the Chat Completions API

Agentic & Coding CLI Usage

Hermes Agent

OpenClaw

OpenCode

Citation

Explore FriendliAI today

README

Pack Summary

Measured Speeds on 1× NVIDIA DGX Spark (GB10, sm_121, 128 GB Unified)

9-Suite (353/356-Item) Downstream Benchmark vs. 4.0 bpw

Fidelity vs. Unquantized ornith-ai/Ornith-1.5-122B BF16 Master

Quickstart: Serving with vLLM 0.32.0 + vcruz305/vllm-exl3

About the Base Model: ornith-ai/Ornith-1.5-122B

Upstream Benchmarks (ornith-ai/Ornith-1.5-122B)

Recommended Sampling & Reasoning Parameters

Using Ornith-1.5-122B via the Chat Completions API

Agentic & Coding CLI Usage

Hermes Agent

OpenClaw

OpenCode

Citation

Measured Speeds on 1× NVIDIA DGX Spark (`GB10`, `sm_121`, 128 GB Unified)

9-Suite (`353`/`356`-Item) Downstream Benchmark vs. `4.0 bpw`

Fidelity vs. Unquantized `ornith-ai/Ornith-1.5-122B` BF16 Master

Quickstart: Serving with `vLLM 0.32.0` + `vcruz305/vllm-exl3`

About the Base Model: `ornith-ai/Ornith-1.5-122B`

Upstream Benchmarks (`ornith-ai/Ornith-1.5-122B`)

Measured Speeds on 1× NVIDIA DGX Spark (`GB10`, `sm_121`, 128 GB Unified)

9-Suite (`353`/`356`-Item) Downstream Benchmark vs. `4.0 bpw`

Fidelity vs. Unquantized `ornith-ai/Ornith-1.5-122B` BF16 Master

Quickstart: Serving with `vLLM 0.32.0` + `vcruz305/vllm-exl3`

About the Base Model: `ornith-ai/Ornith-1.5-122B`

Upstream Benchmarks (`ornith-ai/Ornith-1.5-122B`)