First screen, captured 1 day ago. Hover to scroll the page, or open the video.

Cerebras Deploys Trillion-Parameter AI Models

What HN thinks

Wafer-scale speed meets the eternal question: fast at what, and who actually gets to use it?

Fast but limited practical access
Users say Cerebras flies but only offers weaker models to ordinary developers, and big customers or custom deals get the good capacity.
Architecture curiosity and skepticism
Commenters dig into 2D PE meshes, pipeline parallelism, SRAM limits, KV cache handling, prefill, and failure domains of a 59-wafer pipeline.
Is extreme speed even useful?
Some see huge token rates as a phase change for agents; others argue human review, not speed, is the bottleneck.

Consensus: Most agree the hardware is genuinely fast thanks to on-wafer SRAM, but question its accessibility, cost, and fit against cheaper GPU-served MoE models.

Recommended in the comments

OpenRouter (service): Gives access to stronger models at decent speed when Cerebras' lineup falls short.
DeepSeek V4 Flash (product): Much cheaper to serve at scale than dense Qwen 27B and generally a better model.
Mercury 3 (product): Said to beat Cerebras' Qwen offering on cost and quality at similar speed.

Summarised by Claude from the comment thread 1 day ago.

Points: 349
Comments: 127
Best rank: #2
Posted: 2026-10-25 by gmays
Category: Ai ml

cerebras llm-inference wafer-scale kimi-k3 pipeline-parallelism kv-cache

All captures:

Full page

Captured 1 day ago. Click for the whole screenshot.

Looks similar

Screenshot of baseten.co: Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs

Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs (baseten.co)

215 points by philipkiely 1 year ago | 142 comments

Screenshot of qlabs.sh: NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute

NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute (qlabs.sh)

153 points by sdpmas 7 months ago | 36 comments

Screenshot of baseten.co: The efficient frontier of LLM inference

The efficient frontier of LLM inference (baseten.co)

137 points by philipkiely 1 month ago | 39 comments

Screenshot of cerebras.ai: Cerebras launches Qwen3-235B, achieving 1.5k tokens per second

Cerebras launches Qwen3-235B, achieving 1.5k tokens per second (cerebras.ai)

334 points by mihau 1 year ago | 140 comments

Cerebras press release announcing Qwen3-235B on its Inference Cloud with full 131K context support, billed as the world's fastest frontier AI model.

cerebras qwen3 llm-inference inference-speed

Screenshot of andersmurphy.com: 100k TPS over a billion rows: the unreasonable effectiveness of SQLite

100k TPS over a billion rows: the unreasonable effectiveness of SQLite (andersmurphy.com)

344 points by speckx 10 months ago | 124 comments

Blog post by Anders Murphy benchmarking SQLite at 100k transactions per second over a billion rows, arguing its embedded single-writer design helps it scale.

sqlite database-performance benchmarks clojure