Explore additional models
Frequently
asked questions
No, Qwen3 8B is a dense model that uses all 8.2B parameters on every token. One of Alibaba’s mixture-of-experts options on Crusoe is Qwen3.6 35B A3B, which activates about 3B parameters per token. Dense architecture makes Qwen3 8B a good fit for customization using Serverless Fine-Tuning and high-performance inference via Self-Serve Deployments in Crusoe Intelligence Foundry.
Qwen3.5 9B is the newer generation and the better default for new projects, with a 262,144-token context on Crusoe versus 40,960 for Qwen3 8B, plus improved reasoning and multilingual quality. Qwen3 8B remains the more widely benchmarked base, with a larger library of community fine-tunes and tooling, and is the right pick for short-prompt workloads. Both are available for Serverless Fine-Tuning and Self-Serve Deployments in Crusoe Intelligence Foundry.
Qwen3 8B is billed by how you use it. Serverless Fine-Tuning is billed per 1M training tokens, where training tokens equal dataset tokens multiplied by epochs. Self-Serve Deployments are billed per GPU hour based on the GPU you choose, giving you predictable costs at production volume. Current rates are shown in the pricing section above.
Qwen3 8B is one of the most widely benchmarked small open models, with strong results on reasoning, math, and coding for its size, especially in thinking mode. Its mature ecosystem of community fine-tunes and tooling makes it a reliable base for focused production tasks.
Yes. Use Serverless Fine-Tuning to customize Qwen3 8B on your own data. Upload a JSONL or Parquet dataset, run a LoRA job with pre-configured defaults, then go live in one click using Self-Serve Deployments or download the weights under Apache 2.0. A single fine-tuned checkpoint keeps both thinking and non-thinking modes. Self-Serve Deployments offer inference profiles optimized for throughput or responsiveness. You can also request a Tailored Deployment for bespoke inference optimization and a dedicated endpoint.
In thinking mode, Qwen3 8B reasons step by step inside a thinking block before answering, which improves accuracy on math, code, and multi-step tasks at the cost of latency and output length. In non-thinking mode, it answers directly. Both live in the same checkpoint and are switched per request with the /think and /no_think tags or the enable_thinking parameter.


%204.png)
