Instructions to use Sparset/Thinklite-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Sparset/Thinklite-27B with Transformers:

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

pipe = pipeline("image-text-to-text", model="Sparset/Thinklite-27B")
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("Sparset/Thinklite-27B")
model = AutoModelForMultimodalLM.from_pretrained("Sparset/Thinklite-27B", 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 Sparset/Thinklite-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Sparset/Thinklite-27B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sparset/Thinklite-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Sparset/Thinklite-27B

SGLang

How to use Sparset/Thinklite-27B 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 "Sparset/Thinklite-27B" \
    --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": "Sparset/Thinklite-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "Sparset/Thinklite-27B" \
        --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": "Sparset/Thinklite-27B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Sparset/Thinklite-27B with Docker Model Runner:
```
docker model run hf.co/Sparset/Thinklite-27B
```

Mintzs commited on 7 days ago

Commit

0a77aae

verified ·

1 Parent(s): 3c8e284

Model card: styled tables

Browse files

Files changed (1) hide show

README.md +201 -160

README.md CHANGED Viewed

@@ -1,160 +1,201 @@
----
-license: apache-2.0
-base_model: Qwen/Qwen3.8-27B
-pipeline_tag: image-text-to-text
-library_name: transformers
-tags:
-- thinklite
-- reasoning
-- efficient-reasoning
-- qwen3.8
----
-# Thinklite-27B
-**Qwen3.8-27B that thinks with half the tokens.** At `reasoning_effort="xhigh"`, Thinklite-27B uses 47% less
-reasoning than Qwen3.8-27B on average over six benchmarks and writes half as many output tokens, so answers arrive
-about 2x faster. At `reasoning_effort="low"` it uses 28% less reasoning.
-| Model | xhigh: reasoning vs base | xhigh: speed-up | low: reasoning vs base | low: speed-up |
-|---|---|---|---|---|
-| **Thinklite-27B** | **-47.0%** | **2.0x** | **-28.0%** | **1.3x** |
-| ThinkingCap | -45.8% | 1.7x | -16.4% | 1.2x |
-| Swift 1.5 | -35.6% | 1.6x | -12.7% | 1.1x |
-- *Reasoning vs base*: the change in mean reasoning tokens per answer against Qwen3.8-27B at the same effort,
-  averaged over the six benchmarks below.
-- *Speed-up*: how much faster the full set of answers (thinking plus answer, all 4,291 requests) is generated than
-  with Qwen3.8-27B. All four models share the Qwen3.8-27B architecture and decode at the same speed (Thinklite-27B
-  26.3 vs base 26.2 tokens/s for a single user on the same GPU), so the speed-up is the base model's output tokens
-  divided by the model's.
-Accuracy for every benchmark is in the tables below.
-## How it works
-Reasoning models spend much of their output on thinking that does not change the answer: restating the question,
-repeating a step, re-checking a result that was already checked. Thinklite-27B is fine-tuned from Qwen3.8-27B on its
-own verified-correct answers to reach the same answers by a shorter route. Only the thinking gets shorter; the visible
-answer keeps its full length and format. It is a drop-in replacement for Qwen3.8-27B with the same chat template,
-tokenizer and `reasoning_effort` control.
-## Benchmarks
-**Setup:**
-- We ran all four models ourselves on one frozen suite: identical prompts, chat template, sampling, token caps and
-  scoring. Compared models: [ThinkingCap](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B) and
-  [Swift 1.5](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b), both also built on Qwen3.8-27B.
-- vLLM 0.30.0, bf16 weights, FP8 KV cache, temperature 1.0, top_p 0.95, top_k 20, 64 concurrent requests, one NVIDIA
-  RTX PRO 6000 Blackwell (96 GB). All answers were scored together on one machine.
-- One sample per question (two on AIME 2025). Answers that hit their cap were rerun once with a 131,072-token cap;
-  truncated or unparseable answers count as incorrect.
-- **GSM8K**: 1,319 problems
-- **AIME 2025**: 30 problems x 2 samples, avg@2
-- **MMLU-Pro**: 1,000-question stratified subset
-- **IFEval**: 541 prompts, prompt-level strict
-- **LiveCodeBench v6**: 131 problems (2025-02 to 2025-04), pass@1
-- **BFCL v4**: 1,240 non-live function-calling cases
-### `reasoning_effort="xhigh"`
-Accuracy (%)
-| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
-|---|---|---|---|---|
-| GSM8K | 96.8 | 97.0 | 96.9 | 97.2 |
-| AIME 2025 | 95.0 | 100.0 | 100.0 | 93.3 |
-| MMLU-Pro | 84.4 | 86.5 | 86.6 | 86.3 |
-| IFEval | 89.1 | 90.9 | 90.9 | 91.5 |
-| LiveCodeBench v6 | 84.7 | 84.7 | 84.7 | 84.0 |
-| BFCL v4 | 88.9 | 89.0 | 87.6 | 88.7 |
-| **Average** | **89.82** | **91.37** | **91.12** | **90.17** |
-Reasoning tokens per answer (mean; change vs Qwen3.8-27B)
-| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
-|---|---|---|---|---|
-| GSM8K | 163 (-63%) | 141 (-68%) | 267 (-40%) | 443 |
-| AIME 2025 | 11,236 (-43%) | 13,228 (-33%) | 12,961 (-34%) | 19,642 |
-| MMLU-Pro | 1,343 (-69%) | 1,785 (-58%) | 2,404 (-44%) | 4,295 |
-| IFEval | 1,082 (-41%) | 891 (-51%) | 1,031 (-44%) | 1,830 |
-| LiveCodeBench v6 | 17,079 (-38%) | 21,936 (-20%) | 18,770 (-32%) | 27,589 |
-| BFCL v4 | 143 (-28%) | 113 (-44%) | 159 (-20%) | 200 |
-| **Average change** | **-47.0%** | **-45.8%** | **-35.6%** | - |
-| **All output tokens, 4,291 requests** | **5,822,599 (-50%)** | **6,724,080 (-42%)** | **7,319,314 (-37%)** | **11,533,518** |
-### `reasoning_effort="low"`
-Accuracy (%)
-| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
-|---|---|---|---|---|
-| GSM8K | 96.7 | 97.2 | 97.1 | 97.4 |
-| AIME 2025 | 81.7 | 86.7 | 90.0 | 90.0 |
-| MMLU-Pro | 83.3 | 84.8 | 84.7 | 84.4 |
-| IFEval | 84.3 | 85.8 | 86.0 | 86.1 |
-| LiveCodeBench v6 | 71.0 | 72.5 | 73.3 | 72.5 |
-| BFCL v4 | 91.2 | 91.5 | 91.2 | 90.8 |
-| **Average** | **84.70** | **86.40** | **87.04** | **86.88** |
-Reasoning tokens per answer (mean; change vs Qwen3.8-27B)
-| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
-|---|---|---|---|---|
-| GSM8K | 144 (-21%) | 160 (-12%) | 173 (-5%) | 183 |
-| AIME 2025 | 6,521 (-46%) | 9,324 (-23%) | 9,416 (-22%) | 12,125 |
-| MMLU-Pro | 854 (-23%) | 908 (-18%) | 963 (-13%) | 1,103 |
-| IFEval | 386 (-41%) | 537 (-19%) | 557 (-15%) | 659 |
-| LiveCodeBench v6 | 8,882 (-17%) | 9,183 (-15%) | 9,341 (-13%) | 10,741 |
-| BFCL v4 | 78 (-19%) | 85 (-12%) | 90 (-7%) | 96 |
-| **Average change** | **-28.0%** | **-16.4%** | **-12.7%** | - |
-| **All output tokens, 4,291 requests** | **3,748,880 (-23%)** | **4,121,634 (-16%)** | **4,303,835 (-12%)** | **4,896,249** |
-## Usage
-**vLLM**
-```bash
-vllm serve Sparset/thinklite-27B --reasoning-parser qwen3 --max-model-len 135168 --kv-cache-dtype fp8
-```
-```python
-from openai import OpenAI
-client = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
-reply = client.chat.completions.create(
-    model="Sparset/thinklite-27B",
-    messages=[{"role": "user", "content": "How many primes are there below 100?"}],
-    temperature=1.0, top_p=0.95,
-    extra_body={"top_k": 20, "chat_template_kwargs": {"enable_thinking": True, "reasoning_effort": "xhigh"}},
-)
-print(reply.choices[0].message.content)
-```
-Set `reasoning_effort` to `"low"`, `"medium"` or `"xhigh"`, exactly as with Qwen3.8-27B. The benchmarks used
-temperature 1.0, top_p 0.95 and top_k 20, as recommended for the base model.
-**Transformers**
-```python
-from transformers import AutoModelForImageTextToText, AutoProcessor
-model = AutoModelForImageTextToText.from_pretrained("Sparset/thinklite-27B", dtype="auto", device_map="auto")
-processor = AutoProcessor.from_pretrained("Sparset/thinklite-27B")
-```
-## Files
-Merged bf16 safetensors with the tokenizer, chat template and processor files of Qwen3.8-27B.
-## Limitations
-- One benchmark run per effort; the small benchmarks (AIME 2025 with 60 answers, LiveCodeBench with 131) carry wide
-  confidence intervals.
-- At xhigh, accuracy is below the base model on IFEval (89.1 vs 91.5) and MMLU-Pro (84.4 vs 86.3).
-- At low effort, the average accuracy is 2.2 points below the base model, mostly from AIME 2025 (81.7 vs 90.0).
-- The vision encoder is unchanged from Qwen3.8-27B; training data was text only.
-- No safety tuning beyond the base model's. Check the model on your own tasks before relying on it.
-## License
-Apache-2.0, the license of the base model [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).

+---
+license: apache-2.0
+base_model: Qwen/Qwen3.8-27B
+pipeline_tag: image-text-to-text
+library_name: transformers
+tags:
+- thinklite
+- reasoning
+- efficient-reasoning
+- qwen3.8
+---
+# Thinklite-27B
+**Qwen3.8-27B that thinks with half the tokens.** At `reasoning_effort="xhigh"`, Thinklite-27B uses 47% less
+reasoning than Qwen3.8-27B on average over six benchmarks and writes half as many output tokens, so answers arrive
+about 2x faster. At `reasoning_effort="low"` it uses 28% less reasoning.
+<style>
+.tl-table { width:100%; table-layout:fixed; border-collapse:separate; border-spacing:0; overflow:hidden; border:1px solid #232835; border-radius:16px; background:#0B0D12; font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,sans-serif; font-size:14px; color:#C9CED8; }
+.tl-table th { padding:12px 8px; text-align:center; font-weight:600; color:#8A92A3; background:#181C26; border-right:1px solid #232835; border-bottom:1px solid #232835; }
+.tl-table td { padding:12px 8px; text-align:center; color:#C9CED8; background:#0B0D12; border-right:1px solid #232835; border-bottom:1px solid #232835; vertical-align:middle; font-variant-numeric:tabular-nums; }
+.tl-table tr > :last-child { border-right:0; }
+.tl-table tbody tr:last-child td { border-bottom:0; }
+.tl-table .group { color:#FFFFFF; font-weight:700; border-bottom:2px solid #3B4E8C; }
+.tl-table .name { padding-left:16px; text-align:left; color:#FFFFFF; font-weight:600; }
+.tl-table .tl { background:#151B2E; }
+.tl-table th.tl { color:#7C9CFF; }
+.tl-table td.tl strong, .tl-table td.tl { color:#FFFFFF; }
+.tl-table strong { color:#FFFFFF; }
+.tl-table .up { color:#7C9CFF; font-weight:700; }
+.tl-table .avg td { border-top:2px solid #3B4E8C; font-weight:700; }
+@media (max-width: 640px) {
+  .tl-table { display:block !important; overflow-x:auto !important; table-layout:auto !important; }
+  .tl-table th, .tl-table td { min-width:92px; }
+  .tl-table th:first-child, .tl-table td:first-child { min-width:140px; }
+}
+</style>
+<table class="tl-table">
+  <thead>
+    <tr><th rowspan="2" style="text-align:left;padding-left:16px;vertical-align:bottom;">Model</th><th class="group" colspan="2">reasoning_effort = xhigh</th><th class="group" colspan="2">reasoning_effort = low</th></tr>
+    <tr><th>Reasoning vs base</th><th>Speed-up</th><th>Reasoning vs base</th><th>Speed-up</th></tr>
+  </thead>
+  <tbody>
+    <tr><td class="name tl">Thinklite-27B</td><td class="tl"><strong>&#8595; 47.0%</strong></td><td class="tl"><strong>2.0x</strong></td><td class="tl"><strong>&#8595; 28.0%</strong></td><td class="tl"><strong>1.3x</strong></td></tr>
+    <tr><td class="name">ThinkingCap</td><td>&#8595; 45.8%</td><td>1.7x</td><td>&#8595; 16.4%</td><td>1.2x</td></tr>
+    <tr><td class="name">Swift 1.5</td><td>&#8595; 35.6%</td><td>1.6x</td><td>&#8595; 12.7%</td><td>1.1x</td></tr>
+  </tbody>
+</table>
+- *Reasoning vs base*: the change in mean reasoning tokens per answer against Qwen3.8-27B at the same effort,
+  averaged over the six benchmarks below.
+- *Speed-up*: how much faster the full set of answers (thinking plus answer, all 4,291 requests) is generated than
+  with Qwen3.8-27B. All four models share the Qwen3.8-27B architecture and decode at the same speed (Thinklite-27B
+  26.3 vs base 26.2 tokens/s for a single user on the same GPU), so the speed-up is the base model's output tokens
+  divided by the model's.
+Accuracy for every benchmark is in the tables below.
+## How it works
+Reasoning models spend much of their output on thinking that does not change the answer: restating the question,
+repeating a step, re-checking a result that was already checked. Thinklite-27B is fine-tuned from Qwen3.8-27B on its
+own verified-correct answers to reach the same answers by a shorter route. Only the thinking gets shorter; the visible
+answer keeps its full length and format. It is a drop-in replacement for Qwen3.8-27B with the same chat template,
+tokenizer and `reasoning_effort` control.
+## Benchmarks
+**Setup:**
+- We ran all four models ourselves on one frozen suite: identical prompts, chat template, sampling, token caps and
+  scoring. Compared models: [ThinkingCap](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B) and
+  [Swift 1.5](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27b), both also built on Qwen3.8-27B.
+- vLLM 0.30.0, bf16 weights, FP8 KV cache, temperature 1.0, top_p 0.95, top_k 20, 64 concurrent requests, one NVIDIA
+  RTX PRO 6000 Blackwell (96 GB). All answers were scored together on one machine.
+- One sample per question (two on AIME 2025). Answers that hit their cap were rerun once with a 131,072-token cap;
+  truncated or unparseable answers count as incorrect.
+- **GSM8K**: 1,319 problems
+- **AIME 2025**: 30 problems x 2 samples, avg@2
+- **MMLU-Pro**: 1,000-question stratified subset
+- **IFEval**: 541 prompts, prompt-level strict
+- **LiveCodeBench v6**: 131 problems (2025-02 to 2025-04), pass@1
+- **BFCL v4**: 1,240 non-live function-calling cases
+### `reasoning_effort="xhigh"`
+<table class="tl-table">
+  <thead>
+    <tr><th rowspan="2" style="width:20%;text-align:left;padding-left:16px;vertical-align:bottom;">Benchmark</th><th class="group" colspan="4">Accuracy (%)</th><th class="group" colspan="3">Reasoning tokens vs Qwen3.8-27B</th></tr>
+    <tr><th class="tl">Thinklite-27B</th><th>ThinkingCap</th><th>Swift 1.5</th><th>Qwen3.8-27B</th><th class="tl">Thinklite-27B</th><th>ThinkingCap</th><th>Swift 1.5</th></tr>
+  </thead>
+  <tbody>
+    <tr><td class="name">GSM8K</td><td class="tl">96.8</td><td>97.0</td><td>96.9</td><td><strong>97.2</strong></td><td class="tl">&#8595; 63%</td><td><strong>&#8595; 68%</strong></td><td>&#8595; 40%</td></tr>
+    <tr><td class="name">AIME 2025</td><td class="tl">95.0</td><td><strong>100.0</strong></td><td><strong>100.0</strong></td><td>93.3</td><td class="tl"><strong>&#8595; 43%</strong></td><td>&#8595; 33%</td><td>&#8595; 34%</td></tr>
+    <tr><td class="name">MMLU-Pro</td><td class="tl">84.4</td><td>86.5</td><td><strong>86.6</strong></td><td>86.3</td><td class="tl"><strong>&#8595; 69%</strong></td><td>&#8595; 58%</td><td>&#8595; 44%</td></tr>
+    <tr><td class="name">IFEval</td><td class="tl">89.1</td><td>90.9</td><td>90.9</td><td><strong>91.5</strong></td><td class="tl">&#8595; 41%</td><td><strong>&#8595; 51%</strong></td><td>&#8595; 44%</td></tr>
+    <tr><td class="name">LiveCodeBench v6</td><td class="tl"><strong>84.7</strong></td><td><strong>84.7</strong></td><td><strong>84.7</strong></td><td>84.0</td><td class="tl"><strong>&#8595; 38%</strong></td><td>&#8595; 20%</td><td>&#8595; 32%</td></tr>
+    <tr><td class="name">BFCL v4</td><td class="tl">88.9</td><td><strong>89.0</strong></td><td>87.6</td><td>88.7</td><td class="tl">&#8595; 28%</td><td><strong>&#8595; 44%</strong></td><td>&#8595; 20%</td></tr>
+    <tr class="avg"><td class="name">Average</td><td class="tl">89.82</td><td>91.37</td><td>91.12</td><td>90.17</td><td class="tl">&#8595; 47.0%</td><td>&#8595; 45.8%</td><td>&#8595; 35.6%</td></tr>
+  </tbody>
+</table>
+<details>
+<summary>Mean reasoning tokens per answer</summary>
+| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
+|---|---|---|---|---|
+| GSM8K | 163 (-63%) | 141 (-68%) | 267 (-40%) | 443 |
+| AIME 2025 | 11,236 (-43%) | 13,228 (-33%) | 12,961 (-34%) | 19,642 |
+| MMLU-Pro | 1,343 (-69%) | 1,785 (-58%) | 2,404 (-44%) | 4,295 |
+| IFEval | 1,082 (-41%) | 891 (-51%) | 1,031 (-44%) | 1,830 |
+| LiveCodeBench v6 | 17,079 (-38%) | 21,936 (-20%) | 18,770 (-32%) | 27,589 |
+| BFCL v4 | 143 (-28%) | 113 (-44%) | 159 (-20%) | 200 |
+| **Average change** | **-47.0%** | **-45.8%** | **-35.6%** | - |
+| **All output tokens, 4,291 requests** | **5,822,599 (-50%)** | **6,724,080 (-42%)** | **7,319,314 (-37%)** | **11,533,518** |
+</details>
+### `reasoning_effort="low"`
+<table class="tl-table">
+  <thead>
+    <tr><th rowspan="2" style="width:20%;text-align:left;padding-left:16px;vertical-align:bottom;">Benchmark</th><th class="group" colspan="4">Accuracy (%)</th><th class="group" colspan="3">Reasoning tokens vs Qwen3.8-27B</th></tr>
+    <tr><th class="tl">Thinklite-27B</th><th>ThinkingCap</th><th>Swift 1.5</th><th>Qwen3.8-27B</th><th class="tl">Thinklite-27B</th><th>ThinkingCap</th><th>Swift 1.5</th></tr>
+  </thead>
+  <tbody>
+    <tr><td class="name">GSM8K</td><td class="tl">96.7</td><td>97.2</td><td>97.1</td><td><strong>97.4</strong></td><td class="tl"><strong>&#8595; 21%</strong></td><td>&#8595; 12%</td><td>&#8595; 5%</td></tr>
+    <tr><td class="name">AIME 2025</td><td class="tl">81.7</td><td>86.7</td><td><strong>90.0</strong></td><td><strong>90.0</strong></td><td class="tl"><strong>&#8595; 46%</strong></td><td>&#8595; 23%</td><td>&#8595; 22%</td></tr>
+    <tr><td class="name">MMLU-Pro</td><td class="tl">83.3</td><td><strong>84.8</strong></td><td>84.7</td><td>84.4</td><td class="tl"><strong>&#8595; 23%</strong></td><td>&#8595; 18%</td><td>&#8595; 13%</td></tr>
+    <tr><td class="name">IFEval</td><td class="tl">84.3</td><td>85.8</td><td>86.0</td><td><strong>86.1</strong></td><td class="tl"><strong>&#8595; 41%</strong></td><td>&#8595; 19%</td><td>&#8595; 15%</td></tr>
+    <tr><td class="name">LiveCodeBench v6</td><td class="tl">71.0</td><td>72.5</td><td><strong>73.3</strong></td><td>72.5</td><td class="tl"><strong>&#8595; 17%</strong></td><td>&#8595; 15%</td><td>&#8595; 13%</td></tr>
+    <tr><td class="name">BFCL v4</td><td class="tl">91.2</td><td><strong>91.5</strong></td><td>91.2</td><td>90.8</td><td class="tl"><strong>&#8595; 19%</strong></td><td>&#8595; 12%</td><td>&#8595; 7%</td></tr>
+    <tr class="avg"><td class="name">Average</td><td class="tl">84.70</td><td>86.40</td><td>87.04</td><td>86.88</td><td class="tl">&#8595; 28.0%</td><td>&#8595; 16.4%</td><td>&#8595; 12.7%</td></tr>
+  </tbody>
+</table>
+<details>
+<summary>Mean reasoning tokens per answer</summary>
+| Benchmark | Thinklite-27B | ThinkingCap | Swift 1.5 | Qwen3.8-27B |
+|---|---|---|---|---|
+| GSM8K | 144 (-21%) | 160 (-12%) | 173 (-5%) | 183 |
+| AIME 2025 | 6,521 (-46%) | 9,324 (-23%) | 9,416 (-22%) | 12,125 |
+| MMLU-Pro | 854 (-23%) | 908 (-18%) | 963 (-13%) | 1,103 |
+| IFEval | 386 (-41%) | 537 (-19%) | 557 (-15%) | 659 |
+| LiveCodeBench v6 | 8,882 (-17%) | 9,183 (-15%) | 9,341 (-13%) | 10,741 |
+| BFCL v4 | 78 (-19%) | 85 (-12%) | 90 (-7%) | 96 |
+| **Average change** | **-28.0%** | **-16.4%** | **-12.7%** | - |
+| **All output tokens, 4,291 requests** | **3,748,880 (-23%)** | **4,121,634 (-16%)** | **4,303,835 (-12%)** | **4,896,249** |
+</details>
+## Usage
+**vLLM**
+```bash
+vllm serve Sparset/thinklite-27B --reasoning-parser qwen3 --max-model-len 135168 --kv-cache-dtype fp8
+```
+```python
+from openai import OpenAI
+client = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
+reply = client.chat.completions.create(
+    model="Sparset/thinklite-27B",
+    messages=[{"role": "user", "content": "How many primes are there below 100?"}],
+    temperature=1.0, top_p=0.95,
+    extra_body={"top_k": 20, "chat_template_kwargs": {"enable_thinking": True, "reasoning_effort": "xhigh"}},
+)
+print(reply.choices[0].message.content)
+```
+Set `reasoning_effort` to `"low"`, `"medium"` or `"xhigh"`, exactly as with Qwen3.8-27B. The benchmarks used
+temperature 1.0, top_p 0.95 and top_k 20, as recommended for the base model.
+**Transformers**
+```python
+from transformers import AutoModelForImageTextToText, AutoProcessor
+model = AutoModelForImageTextToText.from_pretrained("Sparset/thinklite-27B", dtype="auto", device_map="auto")
+processor = AutoProcessor.from_pretrained("Sparset/thinklite-27B")
+```
+## Files
+Merged bf16 safetensors with the tokenizer, chat template and processor files of Qwen3.8-27B.
+## Limitations
+- One benchmark run per effort; the small benchmarks (AIME 2025 with 60 answers, LiveCodeBench with 131) carry wide
+  confidence intervals.
+- At xhigh, accuracy is below the base model on IFEval (89.1 vs 91.5) and MMLU-Pro (84.4 vs 86.3).
+- At low effort, the average accuracy is 2.2 points below the base model, mostly from AIME 2025 (81.7 vs 90.0).
+- The vision encoder is unchanged from Qwen3.8-27B; training data was text only.
+- No safety tuning beyond the base model's. Check the model on your own tasks before relying on it.
+## License
+Apache-2.0, the license of the base model [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).