Qwen 3.8 27B Hits Cerebras at 1500 tokens/s with 128k Context

Qwen 3.8 27B Now Available on Cerebras: 1500 tokens/s and 128k Context

Cerebras has quietly added Qwen 3.8 27B to its public inference endpoints, offering developers a new open-weight model with impressive speed: ~1500 tokens per second. The announcement, spotted in the company's updated model catalog, puts the 27-billion-parameter model alongside the existing gpt-oss-120b, which runs at roughly 3000 tokens/s. For teams building AI-powered applications, this is more than a spec bump—it's a signal that high-throughput inference for open models is becoming a mainstream option.

The model is available on both the free trial and pay-as-you-go tiers, subject to rate limits and per-token pricing. On the paid tier, context length extends to 128k tokens, double the 64k offered on the free tier. That's a significant jump for applications requiring long-document processing, multi-turn reasoning, or complex agentic workflows.

Why Throughput Matters More Than Parameter Count

At first glance, it might seem odd that a 27B model runs at half the speed of a 120B model on the same hardware. But as the iMasters analysis points out, throughput on Cerebras depends on how a model maps to the Wafer Scale Engine (WSE) architecture, not just raw parameter size. The WSE's massive on-chip memory and custom interconnect allow certain models to achieve near-linear scaling, while others face bottlenecks in routing or memory bandwidth.

For developers, this means the choice of model isn't just about quality or size—it's about latency and cost per token in production. A 27B model at 1500 tokens/s enables real-time applications like streaming chat, code completion, or interactive agents that would be impractical on slower infrastructure. The fact that Cerebras chose to add Qwen 3.8 27B suggests they see strong demand for mid-sized models that balance capability with speed.

Model Compression Transparency: Unpruned, Quantized Only in Storage

Cerebras has also clarified its stance on model compression, which is crucial for teams that need reproducibility and quality guarantees. In the updated documentation, the company states: "All models served through our public endpoints are the original, unpruned versions." This is a direct response to concerns about pruned models being used without disclosure.

The company uses selective weight-only quantization during storage—meaning weights are stored in partial 16-bit, 8-bit, or 4-bit precision, but sensitive layers remain at full precision and are dequantized on the fly. Activations, attention, and KV cache stay in full precision and unquantized. This approach preserves model quality while reducing memory footprint, which is a pragmatic middle ground between raw FP16 and aggressive quantization.

Notably, Cerebras does not host pruned models on public endpoints. Their research into pruning techniques like REAP (Router-weighted Expert Activation Pruning) is shared on Hugging Face for research purposes, but these are not part of the production API. This transparency is a welcome change in an industry where "compressed" models are sometimes served without full disclosure.

What This Means for the AI Inference Landscape

The addition of Qwen 3.8 27B to Cerebras' catalog is a clear move to expand the ecosystem of open-weight models available at high speed. With gpt-oss-120b and now qwen-3.8-27b, Cerebras is positioning itself as the go-to platform for developers who want open models without sacrificing performance.

This also puts pressure on GPU-based inference providers. While GPUs are versatile, they often struggle to match the raw token throughput of Cerebras' custom silicon for specific workloads. For startups and enterprises alike, the ability to serve a 27B model at 1500 tokens/s could significantly reduce infrastructure costs and improve user experience.

However, there are caveats. The free tier's 64k context limit may be restrictive for some use cases, and rate limits apply. Developers should also evaluate whether the model's quality meets their needs—speed is only one part of the equation. The Qwen 3.8 27B is a strong mid-tier model, but for tasks requiring deep reasoning, larger models like gpt-oss-120b might still be preferable.

Technical Specs at a Glance

• Model: Qwen 3.8 27B
• Model ID: qwen-3.8-27b
• Parameters: 27 billion
• Context (free/paid): 64k / 128k tokens
• Speed: ~1500 tokens/s
• Availability: Free trial and pay-as-you-go tiers
• Compression: Unpruned, selective weight-only quantization for storage

Looking Ahead

The arrival of Qwen 3.8 27B on Cerebras is a reminder that the AI inference race is not just about bigger models—it's about delivering the right model at the right speed. As more open-weight models become available on high-performance hardware, developers will have unprecedented flexibility to build applications that were previously impossible due to latency or cost.

For those already using Cerebras, this is a straightforward addition. For others, it's worth a closer look at what the platform can offer. With 1500 tokens/s and 128k context, Qwen 3.8 27B could be the sweet spot for production workloads that need both speed and context depth.

As the model catalog continues to grow, expect to see more mid-sized models join the lineup. The era of fast, open-weight inference is here, and Cerebras is leading the charge.