A 4-bit MXFP4 quantization of Ornith-1.0-35B, produced with qstream. Only the routed MoE experts (~95% of the weights) are quantized to MXFP4; everything quality-sensitive — including the always-on shared expert — stays BF16, bit-identical to the source.

The original model card follows in full below.


Size	22.9 GB (21.3 GiB weights), down from 70 GB BF16 source (33%, −67%)
Format	compressed-tensors `mxfp4-pack-quantized` — per-expert, FP4 E2M1, group-32 e8m0 scales, symmetric
Base	35B-A3B MoE · 256 experts (top-8) + 1 shared · hybrid attention (30 linear-attention / 10 full-attention) · 40 layers · 262 K context · vision encoder · post-trained on Qwen 3.5 for agentic coding

What is quantized

Component	Precision	Why
Routed experts (`mlp.experts.proj.weight`)	MXFP4 (4-bit)	~95% of the weights, sparsely activated (top-8/256) — the only place worth the size win
Shared expert (`shared_expert.proj.weight`)	BF16	dense, runs on every token — quantizing it cost ~half the PPL loss for <1% size, so kept lossless (see below)
Full + linear (SSM) attention, router gates	BF16	sensitive / per-token — kept lossless from the source
Vision encoder + projector	BF16	image path bit-identical to base
Embeddings, lm_head, norms	BF16	unchanged

The config.json ignore list (720 entries) names every BF16 module, so vLLM never expects packed weights for the shared expert, attention, the Gated-DeltaNet SSM layers, the vision tower, or the routers.

Quality & faithfulness

We report deterministic, reproducible faithfulness metrics rather than noisy small-sample downstream task scores.

Metric	Result	What it shows
WikiText-2 perplexity (test, 296 547 tokens)	7.6949 → 7.7675 (+0.94%)	language modeling intact — a broken quant lands in the hundreds
Routed-expert SQNR (median, MXFP4 vs BF16)	≈ 18.7 dB (16.8–19.4 dB across all 30 720 routed-expert tensors)	reconstruction error is just the unavoidable 4-bit rounding

Why this is enough to trust the checkpoint:

Only the routed-expert FFN weights changed. ~95% of the weights are re-quantized to MXFP4; everything else is bit-identical BF16 to the source. The model is the base model except for 4-bit rounding on the routed-expert GEMMs.
+0.94% perplexity on the full WikiText-2 test set is well inside production-grade territory (a broken quant lands in the hundreds). The sensitive token-mixing (full + linear attention), the shared expert, routing, and vision paths are untouched.
Keeping the shared expert in BF16 matters. Quantizing it too lands at 7.8563 (+2.10%) — i.e. that one dense, always-on FFN (<1% of weights, ~0.2 GB) accounts for ~half the total degradation, because its error hits every token at every layer. Excluding it more than halves the PPL loss for a negligible size cost.
The rounding error is diffuse, not concentrated. Across the 30 720 quantized routed-expert tensors the per-tensor SQNR sits in a tight ~18.7 dB band, and the worst 10% of tensors account for only ~10.6% of total error — no degenerate hotspots, just uniform 4-bit rounding.

PPL is measured over vLLM prompt_logprobs (evals/eval_ppl.py); SQNR via qstream-analyze (weight reconstruction, MXFP4 dequant vs the BF16 source).

Fidelity, footprint & provenance

Vision is untouched: the vision encoder + projector stay BF16 (bit-identical), so image capability equals the base model.
No MTP / speculative decoding: config.json declares mtp_num_hidden_layers: 1, but the base release ships no MTP head weights (confirmed by the upstream authors), so multi-token speculative decoding is not available with this checkpoint. Nothing MTP-related was added here.
Footprint: ~21.3 GiB of weights; fits a single ≥32 GB GPU (KV cache + vision on top).
Provenance: built with qstream @a06edbf from the deepreinforce-ai/Ornith-1.0-35B BF16 release; mixed-precision recipe (routed experts → MXFP4, shared expert + everything else → BF16).

Serving with vLLM

Verified on vllm/vllm-openai:latest-cu130 (vLLM 0.20.0, torch 2.11 + CUDA 13.0). vLLM detects the compressed-tensors config automatically and runs the MXFP4 MoE through its MarlinExperts kernel. On Blackwell this is weight-only FP4 (W4A16): the 4-bit weights are dequantized and the matmul runs in BF16 — full memory savings, lossless activations.

bash

docker run --gpus all --ipc=host -p 8000:8000 \
  vllm/vllm-openai:latest-cu130 \
  --model olka-fi/Ornith-1.0-35B-MXFP4 \
  --gpu-memory-utilization 0.9 \
  --trust-remote-code

The architecture (Qwen3_5MoeForConditionalGeneration) is registered natively in vLLM ≥ 0.20.0 — no fork or runtime patch required.

How it was made

bash

qstream-quantize \
  --model_dir <deepreinforce-ai/Ornith-1.0-35B> \
  --output_dir ./Ornith-1.0-35B-MXFP4 \
  --include_layers '*mlp.experts.*proj.weight' \
  --device cuda --workers 8

The allowlist is routed experts only. The shared expert is intentionally left out so it passes through as BF16 — see the faithfulness note above on why that nearly halves PPL loss.

qstream auto-detects the BF16 source, quantizes only the included routed-expert FFN weights to MXFP4, passes the BF16 remainder through unchanged, and writes the mixed-precision config.json. --include_layers is an allowlist: anything not matched (the shared expert, attention, SSM, vision, routers, embeddings) is guaranteed to pass through untouched.

License

Inherits the MIT license from the base model. This is a derivative (quantized) work of Ornith-1.0-35B.

Original model card

Ornith-1.0-35B

Aloha! 🌺 Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding.

Highlights:

State-of-the-Art Coding Agents: Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE (post-trained on top of Gemma 4 and Qwen 3.5), achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw.
Self-Improving Training Framework: Ornith-1.0 employs RL to learn to generate not only solution rollouts, but also the scallfold that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions.
Licence: MIT licensed, globally accessible, and free from regional limitations.

Ornith 1.0 35B

This model card documents Ornith-1.0-35B, the lightweight member of the Ornith family, designed for efficient single-GPU deployment.

Benchmarks

	Ornith-1.0-35B	Qwen3.5-35B	Qwen3.6-35B	Gemma4-31B	Qwen3.5-397B
Agentic Coding
Terminal-Bench 2.1 _(Terminus-2)	64.2	41.4	52.5	42.1	53.5
Terminal-Bench 2.1 _{(Claude Code)}	62.8	38.9	49.2	-	48.6
SWE-bench Verified	75.6	70	73.4	52	76.4
SWE-bench Pro	50.4	44.6	49.5	35.7	51.6
SWE-bench Multilingual	69.3	60.3	67.2	51.7	69.3
NL2Repo	34.6	20.5	29.4	15.5	36.8
Claw-eval Avg	69.8	65.4	68.7	48.5	70.7
SWE Atlas - QnA	37.1	13.2	15.5	-	20.4
SWE Atlas - RF	29.7	10.2	11.4	-	18.4
SWE Atlas - TW	27.8	9.8	13.3	-	18.5

* Terminal-Bench 2.1 (Terminus-2): We evaluate Terminal-Bench 2.1 using the Harbor/Terminus-2 framework with parser=json, temperature=1.0, top_p=1.0, and a 128K context window. Each run uses a 4-hour timeout with 32 CPU cores and 48GB RAM, and results are averaged over 5 runs. We adjust the Qwen chat template to ensure consistency between training and inference (https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B/blob/main/chat_template.jinja), and modify Harbor to align with vLLM's reasoning_content key.
* Terminal-Bench 2.1 (Claude Code): We evaluate Terminal-Bench 2.1 using Claude Code 2.1.126 with parser=json, temperature=1.0, top_p=1.0, max_new_tokens=131072. Results are averaged over 5 runs. Again, Qwen chat template needs to be modified.
* SWE-Bench Verified, Pro and Multilingual: using OpenHands harness with temp=1.0, top_p=0.95, 256k context window.
* SWE Atlas QnA, RF, TW: using mini SWE agent harness with temp=1.0, top_p=0.95, 128K context window. Results are averaged over 5 runs.
* NL2Repo: with temperature=1.0, top_p=1.0, 400K context, 48K output and anti-hacking filters.
* ClawEval: An agentic code benchmark over real-user task distributions; temp=0.6 and 256K context.

Quickstart

📝 NOTE

Ornith-1.0-35B is a reasoning model: by default the assistant turn opens with a <think> … </think> block before the final answer. The serving recipes below enable a reasoning parser so the chain-of-thought is returned in a separate reasoning_content field, and a tool-call parser so the model's <tool_call> blocks are surfaced as OpenAI-style tool_calls.

Serving Ornith-1.0-35B requires recent runtimes:

Transformers ≥ 5.8.1
vLLM ≥ 0.19.1
SGLang ≥ 0.5.9

Serving Ornith-1.0-35B

The two recipes below stand up an OpenAI-compatible server on a single 8×80GB GPU node (tensor-parallel 8). Adjust --tensor-parallel-size / --tp to the number of GPUs you have.

vLLM

bash

vllm serve deepreinforce-ai/Ornith-1.0-35B \
    --served-model-name Ornith-1.0-35B \
    --tensor-parallel-size 8 \
    --host 0.0.0.0 --port 8000 \
    --max-model-len 262144 \
    --gpu-memory-utilization 0.90 \
    --enable-prefix-caching \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

SGLang

bash

python -m sglang.launch_server \
    --model-path deepreinforce-ai/Ornith-1.0-35B \
    --served-model-name Ornith-1.0-35B \
    --tp 8 \
    --host 0.0.0.0 --port 8000 \
    --context-length 262144 \
    --mem-fraction-static 0.85 \
    --tool-call-parser qwen3_coder \
    --reasoning-parser qwen3

Hugging Face Transformers

For a quick local test (or to script offline generation), load the model directly with Transformers. Make sure you have a recent release installed — see the Transformers installation guide; Ornith-1.0-35B requires transformers >= 5.8.1.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepreinforce-ai/Ornith-1.0-35B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Write a Python function is_prime(n). Keep it short."}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
generated = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
)
output_ids = generated[0][inputs.input_ids.shape[1]:]

# The reply contains a <think> ... </think> reasoning block followed by the answer.
content = tokenizer.decode(output_ids, skip_special_tokens=True)
print(content)

To split the reasoning trace from the final answer, parse on the </think> marker:

python

text = tokenizer.decode(output_ids, skip_special_tokens=True)
if "</think>" in text:
    reasoning, answer = text.split("</think>", 1)
    reasoning = reasoning.replace("<think>", "").strip()
    answer = answer.strip()
else:
    reasoning, answer = "", text.strip()

Using Ornith-1.0-35B via the Chat Completions API

Once a vLLM or SGLang server is running, talk to it with any OpenAI-compatible client.

Basic Usage

python

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",  # any non-empty string works for a local server
)

response = client.chat.completions.create(
    model="Ornith-1.0-35B",
    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].message
# reasoning_content holds the <think> trace; content holds the final answer.
print("reasoning:", getattr(message, "reasoning_content", None))
print("answer:", message.content)

You can also stream tokens, or hand the model tools — Ornith-1.0-35B emits well-formed function calls that the server parses into the standard tool_calls field:

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-1.0-35B",
    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"}

You can point any OpenAI-compatible SDK (Python, Node.js, etc.) or curl at the same /v1/chat/completions endpoint.

Agentic Usage

Ornith-1.0-35B excels in tool-calling and agentic coding capabilities.

Agent Frameworks

Because Ornith-1.0-35B exposes an OpenAI-compatible endpoint with tool calling, it works out of the box with standard agent frameworks. Below is a minimal example that connects Ornith-1.0-35B to tools through an MCP server.

python

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.getenv("OPENAI_BASE_URL", "http://localhost:8000/v1"),
    api_key=os.getenv("OPENAI_API_KEY", "EMPTY"),
)

tools = [
    {
        "type": "function",
        "function": {
            "name": "run_shell",
            "description": "Run a shell command and return its output.",
            "parameters": {
                "type": "object",
                "properties": {
                    "command": {"type": "string", "description": "The command to run"}
                },
                "required": ["command"],
            },
        },
    }
]

messages = [{"role": "user", "content": "List the Python files in the current directory."}]

response = client.chat.completions.create(
    model="deepreinforce-ai/Ornith-1.0-35B",
    messages=messages,
    tools=tools,
    temperature=0.6,
    top_p=0.95,
)
print(response.choices[0].message)

Examples of using Ornith with agent harness:

Hermes Agent

bash

# Hermes talks to any OpenAI-compatible endpoint — point it at your Ornith server.
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"
export MODEL="deepreinforce-ai/Ornith-1.0-35B"

Atomic.chat/ Ollama / llama.cpp

bash

# Both runtimes load a GGUF build of Ornith (publish one at deepreinforce-ai/Ornith-1.0-35B-GGUF).

# llama.cpp — serve an OpenAI-compatible API on port 8000.
llama-server -hf deepreinforce-ai/Ornith-1.0-35B-GGUF --port 8000 -c 262144

# Ollama — pull and chat with the same GGUF straight from Hugging Face.
ollama run hf.co/deepreinforce-ai/Ornith-1.0-35B-GGUF

OpenClaw

bash

# OpenClaw talks to any OpenAI-compatible endpoint — point it at your Ornith server.
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"
export OPENAI_MODEL="deepreinforce-ai/Ornith-1.0-35B"

Unsloth Studio

bash

pip install unsloth

# Load Ornith for fast local inference or fine-tuning (Python):
#   from unsloth import FastLanguageModel
#   model, tokenizer = FastLanguageModel.from_pretrained(
#       "deepreinforce-ai/Ornith-1.0-35B",
#       max_seq_length=262144,
#       load_in_4bit=True,
#   )

OpenHands

bash

pip install openhands-ai

# OpenHands routes through LiteLLM; the "openai/" prefix selects the OpenAI-compatible path.
export LLM_MODEL="openai/deepreinforce-ai/Ornith-1.0-35B"
export LLM_BASE_URL="http://localhost:8000/v1"
export LLM_API_KEY="EMPTY"

# Launch the CLI (or run the official OpenHands Docker image with the same env vars).
openhands

Coding CLIs

Ornith-1.0-35B is optimized for terminal-based coding agents. Point any OpenAI-compatible coding CLI at your Ornith-1.0-35B endpoint (set OPENAI_BASE_URL and OPENAI_API_KEY) to understand large codebases, automate tedious work, and ship faster.

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:8000/v1", "apiKey": "EMPTY" },
#       "models": { "deepreinforce-ai/Ornith-1.0-35B": { "name": "Ornith-1.0-35B" } }
#     }
#   }
# }

opencode

Citation

If you find our work helpful, feel free to give us a cite.

bibtex

@misc{ornith-35b,
    title = {{Ornith-1.0-35B}: Agentic Coding, Open to All},
    url = {https://deep-reinforce.com/ornith_1_0.html},
    author = {{DeepReinforce Team}},
    year = {2026}
}