Qwen 3.8 27B runs at 1500 tokens/s on Cerebras: what changes for those building with AI

Cerebras added Qwen 3.8 27B to its public model catalog, with an announced speed of ~1500 tokens per second. The model joins the already existing gpt-oss-120b, which runs at ~3000 tok/s, and is available both in the free trial and in the pay-as-you-go tier, subject to rate limits and per-token pricing.

For developers who've already fought inference latency, the number stands out not because Qwen 3.8 is the most capable model on the market, but because Cerebras sells exactly one thing: throughput. The company uses its giant wafers (the Wafer Scale Engine) to serve open models at a speed conventional GPUs can't reach. And throughput changes the kind of application you can build.

The specs of what entered the catalog

Cerebras's two public models today look like this:

| Model | Model ID | Parameters | Context (free / paid) | Speed | |---|---|---|---|---| | OpenAI GPT OSS | gpt-oss-120b | 120 billion | 65k / 131k | ~3000 tok/s | | Qwen 3.8 27B | qwen-3.8-27b | 27 billion | 64k / 128k | ~1500 tok/s |

Qwen 3.8 27B is, therefore, the smaller of the two models in parameters, but runs at half the speed of gpt-oss-120b (which makes sense: throughput on Cerebras depends on how the model maps to the hardware, not just raw size). Context in the paid tier is 128k tokens, double the 64k of the free tier.

Open model, unpruned, with quantization only in storage

A point Cerebras makes a point of stating explicitly in its documentation: the models served on public endpoints are the original, unpruned versions. The company even researches pruning techniques like REAP (Router-weighted Expert Activation Pruning) and publishes those models on Hugging Face for the research community, but they do not enter the production API.

On quantization, the documentation details the approach: Cerebras uses selective quantization only in weight storage (part in 16-bit / 8-bit / 4-bit, aligned with industry standard). Sensitive layers stay at full precision with on-the-fly dequantization, and activations, attention, and KV cache remain at full precision, not quantized. Translating that for your case: you get the original model, with its architecture unchanged, with the promise that compression doesn't degrade inference quality.

No, we are committed to serving the original models for all existing endpoints, without modification. We do not alter model architectures via pruning on our hosted portfolio.

This transparency has practical value: when you pin a model by ID in your code, it's good to know that Cerebras commits to not changing the architecture without notice and to offering any pruned versions under distinct names.

What the community has already flagged as caveats

The Hacker News thread brings reactions that matter more than the announcement itself when deciding whether this fits your project.

On real-world use, gardnr (comment) sums up the feel of the speed and raises a critical question for those doing agentic coding:

I used their Coding Plan for a few months. It is genuinely difficult to keep up with the models. The output is so fast. Qwen 3.8 27B is likely one of the strongest models they've hosted so far.

>

-- gardnr

In the same comment, he asks whether Cerebras has already implemented prompt caching, pointing out that without it "it used to get pretty expensive for agentic coding tasks." This is the kind of detail that determines the real cost of an agent that resends a large context at every step: without prompt caching, each iteration pays for the whole context again.

The context limitation also came up. tacone (comment) notes: "the context size they allow for Qwen is just 128k. Still interesting as a specialized sub-agent but not really well suited for long tasks." In other words: great for fast, specialized sub-agents, less suited for very long-context tasks. The same comment notes that the model wasn't yet available on OpenRouter at the time of the post, which limits those who prefer a single model gateway.

There were also reports of onboarding friction: peri-cl (comment) reported a redirect loop when creating an account, and foundfontic (comment) complained about support being concentrated on Discord. These are operational noise, not technical blockers, but they help calibrate expectations for anyone testing it today.

Why 27B at 1500 tok/s matters for the Brazilian developer

The angle here isn't "just another model." It's the combination of a mid-sized open model (27B, a range that runs reasonably well on accessible hardware) with extremely fast inference on a managed endpoint. This opens two concrete paths:

• Fast sub-agents: in a coding agent pipeline, you can use Qwen 3.8 27B via Cerebras for steps that require many short back-and-forths (planning, routing, tools), exactly where accumulated latency kills the experience. The prompt caching caveat, raised in the thread, is what you need to validate before counting on cost.
• Cost-benefit comparison: being open, Qwen 3.8 27B also runs locally. The Cerebras endpoint then becomes a maximum-speed reference point: you can measure whether it's worth keeping slower local inference or paying per token for a service that delivers ~1500 tok/s.

What remains open for anyone adopting it: whether prompt caching is already active (gardnr's comment suggests it was absent in the past), the exact price per token in the paid tier, and how Qwen 3.8 27B compares in quality to gpt-oss-120b for code tasks. Cerebras's documentation points to the model selection guide and the Quickstart as next steps for anyone who wants to make their first API call and measure it against their own workload.