Cerebras adds Qwen 3.8 27B to public API at ~1500 tokens/s
Cerebras's public model catalog now lists Qwen 3.8 27B, a 27 billion parameter open-weight model, alongside the previously listed OpenAI GPT OSS (120 billion parameters). Cerebras's documentation gives Qwen 3.8 27B's throughput as about 1500 tokens per second, with a context window of 64,000 tokens on the free tier and 128,000 tokens on the paid tier; for comparison, GPT OSS is listed at about 3000 tokens per second, with a 65,000/131,000 token context window on the free and paid tiers respectively. The listing drew a fast-moving Hacker News discussion, close to 500 points and 148 comments within about ten hours.
Separately, Cerebras's own documentation on model compression describes its general policy for everything it serves publicly: none of it is pruned. "All models served through our public endpoints are the original, unpruned versions," the documentation states, and the only compression step applied is selective weight-only quantization during storage, keeping weights at 16-bit, 8-bit or 4-bit precision while activations, attention and the key-value cache run in full, unquantized precision; sensitive layers are stored at full precision with on-the-fly dequantization. Cerebras separately publishes pruned versions of models, produced through its own REAP (Router-weighted Expert Activation Pruning) research, on Hugging Face for research use, but says those are not served through its production API.
On the Hacker News thread, several commenters pushed back on how usable the new listing's speed is in practice: they said the public endpoint's rate limit makes sustained workloads such as coding difficult, and that cached input tokens get no price break, which cuts into the usual savings from prompt caching in multi-turn use. Cerebras's model catalog page itself does not show pricing or rate limits; it only notes that use of the models is subject to separate pricing and rate-limit pages.
Key facts
- Cerebras's public model catalog now lists Qwen 3.8 27B, a 27 billion parameter open-weight model, at a documented throughput of about 1500 tokens per second.
- Qwen 3.8 27B's context window is 64,000 tokens on Cerebras's free tier and 128,000 tokens on the paid tier.
- The catalog's other public model, OpenAI's GPT OSS at 120 billion parameters, is listed faster, at about 3000 tokens per second, with a 65,000/131,000 token context window on the free and paid tiers.
- Cerebras says every model on its public endpoints is unpruned, quantizing only the storage of weights (down to 16, 8 or 4 bit precision) while running activations, attention and the key-value cache at full precision.
- On the Hacker News thread about the listing, commenters said the public endpoint's rate limit and its lack of a cached-token price discount blunt the practical benefit of the added speed for sustained workloads such as coding.
Why it matters
Cerebras's public catalog now offers two models: OpenAI's GPT OSS at 120 billion parameters and the newly added Qwen 3.8 27B at 27 billion parameters. Qwen 3.8 27B is listed at about 1500 tokens per second, slower than GPT OSS's roughly 3000 tokens per second, but it is also under a quarter of GPT OSS's size, so the two read as different options on the same platform rather than direct competitors. The listing was popular enough to draw close to 500 points and 148 comments on Hacker News within about ten hours.
Who it affects
Developers already building on Cerebras's public inference endpoints gain a second, smaller model choice alongside GPT OSS. It also matters to anyone comparing fast open-weight inference providers more broadly: the Hacker News discussion around the listing drew commenters weighing Qwen 3.8 27B's raw speed against the practical cost and limits of using the public endpoint for real workloads.
How to use it
Qwen 3.8 27B is available on Cerebras's public endpoints under the model ID qwen-3.8-27b, listed on the same catalog page as GPT OSS (model ID gpt-oss-120b). Its context window is 64,000 tokens on the free tier and 128,000 tokens on the paid tier. The catalog page itself does not show pricing or exact rate limits; Cerebras's documentation points to separate pricing and rate-limit pages and states only that use of the models is subject to both.
How solid is it
The throughput and context figures come from Cerebras's own model catalog documentation, which lists Qwen 3.8 27B and GPT OSS with matching columns for parameters, context and speed. Cerebras's separate compression documentation, describing its policy for everything it serves publicly, says none of it is pruned: weights are quantized only for storage, at 16-bit, 8-bit or 4-bit precision, while activations, attention and the key-value cache stay in full, unquantized precision during computation, and sensitive layers are kept at full precision with on-the-fly dequantization. Cerebras separately publishes pruned versions of models, made through its own REAP (Router-weighted Expert Activation Pruning) research, on Hugging Face, but says those are for research only and are not served through its production API.
Risks and caveats
Several Hacker News commenters said the speed is hard to use in practice on the public endpoint: they described a rate limit that makes sustained workloads such as coding difficult, and pointed out that cached input tokens are billed at the same rate as freshly processed ones, removing one of the usual savings from prompt caching in multi-turn use. One commenter described the billing policy directly, saying input tokens are billed at the standard rate whether served from cache or processed fresh. Cerebras's catalog page does not itself list pricing or rate-limit numbers, so exact terms should be checked on Cerebras's dedicated pages rather than assumed from the listing alone.
“All models served through our public endpoints are the original, unpruned versions.”
— Cerebras, in its documentation on model compression