Instructions to use Michael-Kozu/Deimos-R1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Michael-Kozu/Deimos-R1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Michael-Kozu/Deimos-R1")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Michael-Kozu/Deimos-R1")
model = AutoModelForMultimodalLM.from_pretrained("Michael-Kozu/Deimos-R1", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Michael-Kozu/Deimos-R1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Michael-Kozu/Deimos-R1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Michael-Kozu/Deimos-R1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Michael-Kozu/Deimos-R1

SGLang

How to use Michael-Kozu/Deimos-R1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Michael-Kozu/Deimos-R1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Michael-Kozu/Deimos-R1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Michael-Kozu/Deimos-R1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Michael-Kozu/Deimos-R1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Michael-Kozu/Deimos-R1 with Docker Model Runner:
```
docker model run hf.co/Michael-Kozu/Deimos-R1
```

Michael-Kozu commited on Jul 16

Commit

0fb1712

verified ·

1 Parent(s): d1cb33f

Final benchmarks: full-budget retest + token-efficiency analysis

Browse files

Files changed (1) hide show

README.md +46 -31

README.md CHANGED Viewed

@@ -22,39 +22,54 @@ The result: **large gains under a constrained thinking budget** — math reasoni
 extraction, and instruction-following all improve over the identically-measured base — from a
 single epoch (~3 GPU-hours on one GB10).
-## Benchmarks — read this first
-All numbers below use a **4,096-token generation budget**. This is a deliberate
-constrained-inference protocol, and it suppresses thinking-mode models severely: Qwen reports
-**79.1 MMLU-Pro / 89.8 IFEval** for this same base model at its recommended settings (32k–81k
-output budget, tuned sampling) — under our 4k budget the base measures 39.4/24.7. **These tables
-therefore show relative performance under a tight token budget, NOT a claim that Deimos R1
-exceeds the base's published full-budget scores** (that comparison has not been run). What the
-deltas demonstrate: Deimos R1 reaches correct answers in far fewer thinking tokens than the base —
-useful precisely when inference budget matters.
-External suite (lm-eval 0.4.12; identical protocol for base and fine-tune: chat template applied,
-few-shot as multi-turn, 4,096-token generation budget, seed 42, gsm8k/IFEval n=150, MMLU-Pro
-25/subtask):
-| Benchmark | Qwen3.5-4B (base) | **Deimos R1** | Δ |
 |---|---|---|---|
-| gsm8k (flexible-extract) | 0.660 | **0.933** | **+0.273** |
-| MMLU-Pro (custom-extract) | 0.394 | **0.634** | **+0.240** |
 | IFEval prompt-strict | 0.247 | **0.320** | +0.073 |
-| IFEval prompt-loose | 0.260 | **0.347** | +0.087 |
-| IFEval inst-strict | 0.433 | **0.479** | +0.046 |
-| IFEval inst-loose | 0.441 | **0.500** | +0.059 |
-| gsm8k (strict `####` match) | 0.673 | 0.233 | −0.440 * |
-\* The one honest regression, and a narrow one: gsm8k-strict rewards *imitating the few-shot
-examples' `#### N` ending* (no explicit instruction is given). SFT gives the model a confident
-answer style that overrides silent example-imitation. Capability is unaffected (0.933 flexible on
-the same outputs) and **instructed** format-following improved across the board (IFEval ↑, MMLU-Pro
-extraction ↑). See "Format adherence" below.
-Internal harness (held-out, contamination-filtered math problems with verified ground truth, n=200):
-reasoning accuracy **0.175 → 0.465 (2.7×)**; IF checks held at 0.917.
 ## Training data

 extraction, and instruction-following all improve over the identically-measured base — from a
 single epoch (~3 GPU-hours on one GB10).
+## Benchmarks — capability per token
+Deimos R1's headline property, measured on an independent open harness (lm-eval 0.4.12, identical
+settings for both models in every comparison): **it matches or beats its base on capability metrics
+while thinking 3–5× fewer tokens.**
+**Full-budget comparison** (32,768-token generation budget, Qwen-recommended sampling —
+temperature 1.0, top_p 0.95, presence_penalty 1.5, thinking on; gsm8k/IFEval n=150, MMLU-Pro
+25/subtask, seed 42):
+| Benchmark | Qwen3.5-4B (base) | **Deimos R1** | Δ | base think-tokens | **R1 think-tokens** |
+|---|---|---|---|---|---|
+| gsm8k (flexible) | 0.860 | **0.907** | +0.047 | 1,778 | **357 (5.0× fewer)** |
+| MMLU-Pro | 0.363 | **0.551** | +0.188 | 1,984 | **677 (2.9× fewer)** |
+| IFEval prompt-loose | 0.260 | **0.353** | +0.093 | 4,808 | **1,189 (4.0× fewer)** |
+| IFEval inst-loose | 0.437 | **0.487** | +0.050 | ″ | ″ |
+| IFEval prompt-strict | 0.260 | 0.267 | +0.007 | ″ | ″ |
+| IFEval inst-strict | 0.437 | 0.429 | −0.008 | ″ | ″ |
+| gsm8k (strict `####`) | 0.727 | 0.333 | −0.394 * | ″ | ″ |
+**Constrained-budget comparison** (4,096-token budget, greedy decoding — the regime where thinking
+economy dominates):
+| Benchmark | base | **Deimos R1** | Δ |
 |---|---|---|---|
+| gsm8k (flexible) | 0.660 | **0.933** | +0.273 |
+| MMLU-Pro | 0.394 | **0.634** | +0.240 |
 | IFEval prompt-strict | 0.247 | **0.320** | +0.073 |
+### Protocol honesty notes
+- **Published-number comparability:** Qwen reports 79.1 MMLU-Pro / 89.8 IFEval for this base using
+  its internal evaluation stack. Under lm-eval's prompts/extraction, the *base itself* measures far
+  below those numbers at any budget (MMLU-Pro 0.36–0.39) — so no row above is comparable to
+  published figures, in either direction. All claims here are strictly like-for-like: same harness,
+  same settings, same seed, both models.
+- IFEval under lm-eval scores the full response including the `<think>` block, which penalizes
+  long thinkers on constraint checks; this affects both models symmetrically (and is partly why
+  short thinking helps here).
+- The two budget tables also differ in decoding (greedy vs Qwen-recommended sampling); compare
+  within a table, not across.
+- \* gsm8k-strict rewards imitating the few-shot examples' `#### N` ending (no explicit
+  instruction). Deimos R1 favors its trained answer style over silent example-imitation — a known,
+  documented trade (see Format adherence). Ask for a format in words and it complies (IFEval ↑,
+  MMLU-Pro extraction ↑).
+Internal harness (held-out, contamination-filtered math with verified ground truth, n=200):
+reasoning accuracy **0.175 → 0.465 (2.7×)** vs base; IF checks held at 0.917.
 ## Training data