DeepSeek R1 Distill Llama 8B
DeepSeek's R1 distill on a Llama 3.1 8B base. It writes out a chain of thought before answering and is the only Llama option among the smaller R1 distills.
Overview
DeepSeek built this model from Llama-3.1-8B, Meta's pretrained base checkpoint without instruction tuning, and fine-tuned it on reasoning samples written by DeepSeek-R1. As with the Qwen distills, training stopped after supervised fine-tuning, with no reinforcement learning stage.
The scores DeepSeek reports for it: 50.4 pass@1 on AIME 2024, 89.1 on MATH-500, 49.0 on GPQA Diamond, 39.6 on LiveCodeBench and 1205 on Codeforces.
Choosing a variant: the closest sibling is DeepSeek R1 Distill Qwen 7B, which starts from a math-tuned Qwen base. Compared with it, this model gives up some math (50.4 vs 55.5 on AIME 2024, 89.1 vs 92.8 on MATH-500) and gains a little on code (39.6 vs 37.6 on LiveCodeBench). GPQA Diamond is a near tie. Licensing also differs: this model's base comes under the Llama 3.1 license, while the Qwen distills start from Apache 2.0 bases. DeepSeek R1 Distill Qwen 1.5B goes smaller still. For more reasoning depth, look at DeepSeek R1 Distill Qwen 14B and DeepSeek R1 Distill Qwen 32B.
What you can build
- Reasoning features in products that already run Llama-family models and want to stay on that lineage.
- Light code reasoning, such as explaining why a short function fails, where it edges out the Qwen 7B distill.
- Side-by-side tests against the Qwen 7B distill, to see how the base model changes the reasoning.
Inputs
Only a prompt is required. Since DeepSeek tells users of its R1 models to avoid a system prompt, write any rules, role or output format straight into the prompt. The optional system prompt box has a default message you can empty.
Conversations need two identifiers on every turn, one for the user and one for the session. With both set, Wiro remembers earlier messages in that session.
Temperature starts at 0.7 (range 0 to 5) and top-p at 0.95. You can also adjust top-k and the repetition and length penalties, bound output length, add stop sequences, fix a seed, and turn sampling or quantization off; both start on.
Outputs
The reply is text. Before the answer, the model writes a reasoning section inside think tags, and Wiro delivers that section as a thinking segment kept apart from the answer. Your app decides whether users see it.
Recommended settings
- Set temperature to 0.6, DeepSeek's recommendation. Its stated range for R1 models is 0.5 to 0.7, chosen to prevent endless repetition.
- Top-p can stay at its 0.95 default, the value DeepSeek evaluated with.
- For math, ask for worked steps and a boxed final answer; DeepSeek uses \boxed{} for this.
- Average several runs before drawing conclusions about quality. DeepSeek recommends this when evaluating R1 models.
Limitations
- On pure math it trails the Qwen 7B distill.
- DeepSeek has observed R1-series models skipping their thinking on some queries, with lower-quality answers as a result.
- R1 targets Chinese and English, according to DeepSeek, and can switch languages mid-reply on other inputs.
- Few-shot examples reduce R1's performance, DeepSeek reports.
Safety & compliance
DeepSeek's release is MIT-licensed, and the model is derived from Llama 3.1 8B Base, which comes under the Llama 3.1 license. Check both sets of terms before shipping.
API quick start
Run DeepSeek-R1-Distill-Llama-8B with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/deepseek-ai/deepseek-r1-distill-llama-8b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



