{"slug": "qwen-3-8-27b-on-cerebras-1500-tokens-per-second", "title": "Qwen 3.8 27B on Cerebras: 1,500 Tokens Per Second", "summary": "Cerebras added Alibaba's Qwen 3.8 27B to its public API, delivering 1,500 tokens per second, enabling a 300-word response in under half a second. The model scores 61.7% on SWE-bench Pro, ahead of Claude Opus 4.6 Max at 53.4%, and is priced at $0.99 per million input tokens and $0.49 per million output tokens, but lacks cache discounts, making it costlier for context-heavy workloads.", "body_md": "Cerebras just added Alibaba’s Qwen 3.8 27B to its public API, and the throughput number is not a benchmark to file away — it’s 1,500 tokens per second. At that speed, a 300-word streaming response completes in under half a second. Multi-step agent loops that previously had to run in the background now finish inside an active user session. The “AI feels slow” problem, for certain workloads, is gone.\n\n## What Qwen 3.8 27B Is\n\nAlibaba released [Qwen 3.8 27B on August 14 under the Apache 2.0 license](https://huggingface.co/Qwen/Qwen3.8-27B). It is a 27-billion-parameter dense multimodal model that accepts text and images, supports tool calling and structured outputs, and uses a Hybrid Gated DeltaNet attention mechanism in which three of every four attention sublayers operate in linear O(n) time — meaning it handles long contexts without the quadratic memory blowup that plagues standard transformers.\n\nThe benchmark story is solid. On SWE-bench Pro, Qwen 3.8 27B scores 61.7 percent, ahead of Claude Opus 4.6 Max at 53.4 percent. On [Artificial Analysis’s composite intelligence index](https://artificialanalysis.ai/models/qwen3-8-27b) it sits at 41 out of 100 — not a frontier model, but a capable one that runs on 17 GB of VRAM at 4-bit quantization if you want to self-host. The native context window is 262K tokens; Cerebras caps it at 128K on the paid tier.\n\n## The Speed Advantage\n\nCerebras achieves this throughput through its [Wafer-Scale Engine](https://www.cerebras.ai/chip). Where GPU inference relies on off-chip High Bandwidth Memory, Cerebras stores model weights in 44 GB of on-chip SRAM on a single 46,225 mm² silicon die. There is no off-chip memory bottleneck. For low-batch, per-user inference — the exact regime where GPUs idle — the WSE architecture runs fast.\n\nTypical GPU inference for a 27B model runs at 60–200 tokens per second per user. Cerebras delivers 1,500. That is not a marginal improvement; it is a different product category.\n\n## Pricing: Cheaper Than It Looks, Until It Isn’t\n\nDeveloper tier pricing is /bin/bash.99 per million input tokens and .49 per million output tokens. A free tier provides one million tokens per day with a 64K context window and five requests per minute. The paid developer tier raises that to 300 requests per minute, a 128K context window, and a 40,960 token maximum output.\n\n| Model | Provider | Input ($/M) | Output ($/M) | \n|---|---|---|---|\n| Qwen 3.8 27B | Cerebras | /bin/bash.99 | .49 | \n| Gemini 3.8 Flash |  | /bin/bash.75 | .75 | \n| GPT-5.6 Sol | OpenAI | .00 | 0.00 | \n\nHere is the catch: Cerebras offers no cache discount. Anthropic gives 75 percent off on cache reads. OpenAI gives 50 percent. Cerebras bills every token at full price, whether it was served from cache or processed fresh. For workloads that resend a large system prompt or growing conversation history on every turn, this collapses the economics quickly. One documented real-world test showed a five-minute coding session costing .60 on Cerebras versus /bin/bash.29 on a cached frontier alternative — 5.6 times more expensive despite being 2.8 times faster in wall-clock time. Speed does not automatically mean cheaper.\n\n## When to Use It and When to Skip It\n\nCerebras Qwen 3.8 27B is the right choice for workloads that are latency-sensitive and context-light: single-turn classification, interactive chat with short histories, rapid tool calls, and real-time document analysis where you process once and move on. At 1,500 tokens per second, the response is back before the user blinks.\n\nIt is the wrong choice for long agentic coding sessions, sustained pipelines that resend large context on every step, or any workload where the 128K context ceiling will bind. And there is an important production caveat: Cerebras’s shared tier carries no uptime SLA. The company replaced Gemma 4 31B on the same tier without notice. If your application needs guaranteed availability, plan accordingly.\n\n## Getting Started\n\nThe Cerebras API is OpenAI-compatible. Swap the base URL and key:\n\n``` python\nfrom openai import OpenAI\n\nclient = OpenAI(\n    api_key=\"your-cerebras-key\",\n    base_url=\"https://api.cerebras.ai/v1\"\n)\n\nresponse = client.chat.completions.create(\n    model=\"qwen-3.8-27b\",\n    messages=[{\"role\": \"user\", \"content\": \"Your prompt here\"}],\n    stream=True\n)\n```\n\nFree tier access is available at [cloud.cerebras.ai](https://cloud.cerebras.ai) with no credit card required for the first million tokens per day. The developer tier is pay-as-you-go.\n\n## The Bigger Picture\n\nThe inference speed war is real, and 2026 has been its most consequential year. Cerebras has held the “fastest LLM API” position for over a year; Groq trails at roughly 800–1,000 tokens per second for comparable model sizes. GPU-based inference providers are closing the gap, but not there yet.\n\nWhat Qwen 3.8 27B on Cerebras represents is a specific intersection: an open-weight model with genuine capability, served at a speed tier that enables new UX patterns, at a price point that undercuts frontier models for the right workloads. It is not a replacement for Claude Fable 5.1 or GPT-6 Astra. It is a tool for a specific job, and for that job, it is currently the best option available.", "url": "https://wpnews.pro/news/qwen-3-8-27b-on-cerebras-1500-tokens-per-second", "canonical_source": "https://byteiota.com/qwen-3-8-27b-on-cerebras-1500-tokens-per-second/", "published_at": "2026-09-07 11:15:54+00:00", "updated_at": "2026-09-07 11:26:33.489690+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-infrastructure", "ai-products"], "entities": ["Cerebras", "Alibaba", "Qwen 3.8 27B", "Claude Opus 4.6 Max", "Artificial Analysis", "Wafer-Scale Engine"], "alternates": {"html": "https://wpnews.pro/news/qwen-3-8-27b-on-cerebras-1500-tokens-per-second", "markdown": "https://wpnews.pro/news/qwen-3-8-27b-on-cerebras-1500-tokens-per-second.md", "text": "https://wpnews.pro/news/qwen-3-8-27b-on-cerebras-1500-tokens-per-second.txt", "jsonld": "https://wpnews.pro/news/qwen-3-8-27b-on-cerebras-1500-tokens-per-second.jsonld"}}