The Pulse
Cerebras Adds Qwen3.8-27B at 1,500 Tokens per Second
Cerebras has added Alibaba’s Qwen3.8-27B to its public inference service, with the model listed at roughly 1,500 output tokens per second. The model is available through Cerebras’ pay-as-you-go developer tier with a 150,000-token-per-minute
AI.info Team ·
Cerebras is putting Alibaba’s Qwen3.8-27B on the same public inference service that it uses to promote some of the fastest open-model response times available through an API. The move gives developers access to a mid-sized Qwen model at a listed speed of roughly 1,500 output tokens per second, but the public tier also imposes limits that may shape which workloads can use it in production.
Cerebras’ support documentation lists Qwen 3.8 27B as one of two models available through its Developer Tier, alongside OpenAI’s GPT OSS 120B. The company’s pay-as-you-go system charges customers for input and output tokens, while Qwen3.8-27B receives a lower token allowance than GPT OSS 120B: 150,000 uncached tokens per minute and 450 requests per minute.
The addition brings one of Alibaba’s newer open-weight models to Cerebras’ wafer-scale inference hardware. It also makes speed, rather than model size alone, the main selling point. Developers can now test Qwen3.8-27B without running the model locally or arranging dedicated hardware, though teams with sustained high-volume traffic will need to seek higher limits and reserved capacity from Cerebras.
Qwen3.8-27B arrives with a 150,000-token ceiling
Cerebras’ Developer Tier FAQ identifies Qwen 3.8 27B as an available pay-as-you-go model and gives it a limit of 150,000 uncached tokens per minute. The same table lists 450 requests per minute. Those limits apply to the self-serve tier and can change, according to the company’s documentation.
The distinction between cached and uncached tokens matters for applications that repeatedly send the same prompt prefix. Cerebras’ API supports prompt caching, which can reduce repeated processing for shared prefixes such as system instructions, retrieval templates or agent context. The published limit, however, is stated in uncached tokens, so developers planning large document or agent workloads still need to measure actual request patterns rather than relying only on the headline throughput.
Qwen3.8-27B is not presented as an enterprise-only deployment. Cerebras says users can purchase credits, consume them based on token usage and configure automatic recharging. Credits expire one year after purchase. If an account reaches a zero balance, the company says its rate limits revert to zero until more credits are added.
Public-tier limits: Cerebras lists Qwen 3.8 27B at 150,000 uncached tokens per minute and 450 requests per minute. Higher limits, service-level agreements and dedicated capacity require a separate sales arrangement.
Why 1,500 tokens per second matters
Cerebras lists Qwen3.8-27B at approximately 1,500 tokens per second on its public inference infrastructure, according to documentation reviewed after the model appeared in the company’s public catalog. The figure describes generation speed under Cerebras’ measurement conditions; actual performance varies with prompt length, concurrency, sampling settings, request shape and account tier.
At that rate, a model can produce long answers or code suggestions in a fraction of the time associated with many conventional GPU-hosted endpoints. Speed matters most in applications that generate substantial output: coding agents, research assistants, document transformation pipelines and systems that make several model calls before returning an answer.
Faster generation does not eliminate other bottlenecks. A long prompt still has to be processed, network latency still affects the first response token and rate limits can constrain parallel requests. A team running hundreds of concurrent agents may exhaust the 150,000-token-per-minute allowance even when each individual response arrives quickly.
That tension gives the Qwen3.8-27B release a more specific value proposition than simply adding another model to a catalog. Cerebras is offering a fast dense model for interactive applications, but the public tier is sized for experimentation and moderate workloads rather than unrestricted deployment at scale.
Cerebras’ public catalog narrows to two headline models
The Qwen addition also changes the composition of Cerebras’ self-serve offering. The company’s support page names Qwen 3.8 27B and GPT OSS 120B as the models currently available through the pay-as-you-go Developer Tier. GPT OSS 120B receives a higher allowance of 500,000 uncached tokens per minute and 1,000 requests per minute.
That contrast reflects different positioning. GPT OSS 120B is the larger model and is marketed for reasoning, coding and agent workflows. Qwen3.8-27B offers a smaller parameter count and lower public-tier capacity, but its roughly 1,500-token-per-second figure gives developers a fast option when response time and operating cost matter more than maximum model scale.
Cerebras has historically used its inference service to host open models from several companies, including Qwen, Meta, Google and OpenAI. Its pay-per-token announcement said the service was designed to make open-weight models available without contracts or dedicated infrastructure. The company also offers subscription products for coding and separate enterprise arrangements with priority routing and dedicated support.
“AI has moved from being a novelty to being useful and productive,” Andrew Feldman, co-founder and chief executive officer of Cerebras Systems, said in the company’s June 23, 2026 earnings announcement. Feldman added that “fast AI is more valuable than slow AI because it is more productive.”
Qwen3.8-27B fits that argument directly. Its appeal is not only the underlying model’s capability; it is the combination of an open-weight model, an OpenAI-compatible API workflow and unusually high generation speed on Cerebras hardware.
The model is open, but the service is not unlimited
Developers can call the model through Cerebras’ inference API rather than managing weights, memory allocation, quantization and serving infrastructure themselves. The company’s public model documentation describes its endpoints as supporting common API features such as streaming and model retrieval, while Cerebras’ public materials explain that customers can begin with a free account and move to paid usage.
Pricing is separate from the rate-limit question. Cerebras says Developer Tier billing is based on input and output token consumption, with different models carrying different prices. The company’s documentation instructs users to consult its current pricing page instead of embedding a fixed price in the support FAQ, which means developers should verify the amount shown in their account before budgeting a deployment.
The arrangement also leaves room between a public endpoint and a dedicated production service. Cerebras says customers that need higher limits, service-level agreements or dedicated capacity should contact its sales team. That separation is familiar in hosted inference, but it matters here because a 1,500-token-per-second benchmark can create demand faster than a public quota can absorb it.
Teams evaluating Qwen3.8-27B should therefore test four things separately: time to first token, sustained output speed, behavior under concurrent load and the number of tokens consumed per completed task. A coding agent that produces a correct patch in one fast pass may be cheaper than a slower model that requires several retries, but the calculation depends on model quality and workflow design rather than raw tokens per second alone.
Alibaba’s model meets Cerebras’ speed-first strategy
Alibaba’s Qwen family has become a common choice for developers seeking open-weight models that can be downloaded, modified or served through third-party providers. Cerebras’ addition gives those developers another way to access Qwen3.8-27B without buying hardware capable of holding the model and its context window.
The release also extends Cerebras’ effort to make inference speed the center of its cloud offering. Earlier company announcements promoted Qwen3 models, GPT OSS and other open models through pay-per-token access. Cerebras has separately announced large infrastructure partnerships, including work with Amazon Web Services to combine AWS Trainium systems with Cerebras CS-3 systems for inference workloads.
Qwen3.8-27B is a smaller and more accessible test of that strategy than a trillion-parameter model. If developers use it for extraction, code repair, classification, summarization or multi-step agents, the practical comparison will be against other fast hosted endpoints and against self-hosting on GPUs—not only against larger frontier models.
For now, the concrete offer is straightforward: Qwen3.8-27B is available through Cerebras’ pay-as-you-go developer service, the company lists approximately 1,500 output tokens per second, and the public quota is 150,000 uncached tokens per minute with 450 requests per minute. Those numbers define the release more clearly than the model announcement itself. They show a service built for developers who want Qwen’s open weights and Cerebras’ speed, but who can operate within a bounded public tier.