Cerebras adds Qwen 3.8 27B at a listed 1,850 tokens per second

Cerebras has started serving Alibaba’s Qwen 3.8 27B through its wafer-scale inference cloud. The open-weight model accepts text and images through an OpenAI-compatible API, with listed generation throughput of approximately 1,850 tokens per second on the shared endpoint. That speed could shorten coding, research, and tool-calling loops that generate long responses.

Cerebras cites a score of 34 on the Artificial Analysis Intelligence Index, a composite benchmark covering reasoning, knowledge, coding, and related tasks. The cited snapshot places the model near GPT-5.6 Luna, DeepSeek V4 Pro, and Claude Sonnet 4.6. Benchmark rankings offer a broad quality signal, while application-specific evaluations remain necessary for tool selection, code generation, vision, and structured output.

The endpoint, line by line

The Cerebras model docs list the following capabilities and limits:

Calls default to high reasoning, which can add internal reasoning tokens before the final answer and increase both latency and usage. Developers using the endpoint as a low-latency chat backend can set reasoning_effort to none. Prompt caching may reduce processing time, while cached prefixes retain the standard input-token price on the shared tier.

Dense weights meet a wafer-sized chip

Qwen 3.8 27B uses a dense architecture, meaning each token follows the same full compute path through the model. Mixture-of-experts models such as Qwen3-235B route tokens through selected expert blocks. The dense design gives the 27B model a more uniform per-token resource profile, which is useful when estimating latency and capacity for repeated agent calls.

Cerebras maps inference onto a wafer-scale processor designed to keep model computation and memory access on one large device. Conventional GPU clusters exchange data across multiple accelerator packages, adding communication overhead during generation. Cerebras used the same architectural argument for its earlier Qwen3-235B launch, where the company claimed 30 times the speed and one-tenth the cost of selected closed-model alternatives.

A 2,000-token completion would require about 1.1 seconds of generation at 1,850 tokens per second, compared with 40 seconds at 50 tokens per second. End-to-end application latency also includes queueing, network transit, prompt processing, time to first token, and tool execution. Teams should benchmark complete workflows rather than infer user-facing latency from decode throughput alone.

Wire it up in one request

Existing OpenAI-compatible clients can use the Cerebras base URL and the new model ID. A minimal streaming request looks like this:

curl https://api.cerebras.ai/v1/chat/completions \
  -H "Authorization: Bearer $CEREBRAS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-3.8-27b",
    "messages": [
      {
        "role": "user",
        "content": "Review this patch and list any correctness risks."
      }
    ],
    "reasoning_effort": "none",
    "stream": true
  }'

The example disables reasoning for a latency-sensitive interaction; omitting that field restores the high default. Vision requests can include base64-encoded PNG or JPEG content, while agent workflows can combine images, parallel tool calls, and strict JSON output in the same API flow.

Shared capacity has hard edges

Cerebras has described its shared pay-as-you-go service as an experimentation tier and directs production workloads toward enterprise or dedicated deployments. The shared service carries no production SLA, and Cerebras has previously changed its available model roster without a migration window. Applications that depend on model pinning or guaranteed capacity need stronger deployment terms.

The 150K tokens-per-minute limit also constrains concurrency. A stream sustaining 1,850 generated tokens per second would produce about 111K tokens in one minute, leaving limited aggregate capacity for prompts and simultaneous requests. The 128K paid context window can further raise costs in stateless agent loops because clients commonly resend conversation history and tool results on each turn.

Production systems that require stable latency, reserved throughput, or predictable model availability should use an enterprise agreement or dedicated deployment. The shared endpoint remains suitable for prototypes, evaluations, burst workloads, and experiments where generation speed materially affects the workflow.

27B fits bounded agent work

Alibaba’s Qwen 3.8 lineup spans dense and mixture-of-experts designs, including the 27B model, a 235B MoE flagship, and a Flash variant. The hosted 27B endpoint targets tasks that benefit from low decode latency while retaining vision, reasoning, structured output, and tool use.

Downloadable weights also give teams the option to fine-tune or self-host the model under its license. Its 27B parameter count reduces storage requirements relative to a 235B checkpoint, although long contexts still consume substantial accelerator memory through the key-value cache, with exact requirements depending on quantization, batch size, and serving software.

Teams already routing Qwen3-235B traffic through Cerebras can evaluate the 27B model on bounded subtasks such as classification, retrieval synthesis, tool selection, patch drafting, and image-assisted analysis. Useful measurements include task success, time to first token, end-to-end latency, and total token consumption with reasoning enabled and disabled. The endpoint is available through Cerebras Cloud.