Cerebras

Key Terms

• Wafer-Scale Engine (WSE-3) - Cerebras's chip architecture delivering up to ~3,000 tokens/sec on GPT OSS 120B, the core differentiator versus GPU inference. Source: Inference-Docs – Overview
• Dedicated Endpoints - enterprise reserved capacity with 30+ additional model families, production SLAs, custom weights, and fine-tuning; the only way to run more than the two shared-API models. Source: Inference-Docs – Overview
• REAP - Router-weighted Expert Activation Pruning, Cerebras's MoE compression research; pruned models are published on HuggingFace for research only and explicitly not served on the public API. Source: Inference-Docs – Overview

Latest Changes

Changes since the 2026-07 report.

July watch items verified:

• GLM 4.7 deprecation (Aug 17): confirmed. The shared model catalog now lists exactly two models: GPT OSS 120B and Gemma 4 31B. No replacement was added. Source: Inference-Docs – Overview
• Cerebras Code subscription reopening: failed (4th consecutive month). The pricing page no longer lists the Cerebras Code Pro/Max subscription product at all; only Inference API tiers (Free Trial $5 credits, Developer, Enterprise) are shown. Source: Cerebras – Pricing
• AMD partnership availability (H2 2026): still-pending. No update found on the July-announced AMD Helios + WSE disaggregated inference pairing. Source: Cerebras – Partners

New August changes:

• Infrastructure: Cerebras and Compute Nordic Finland announced a 165 MW AI data centre in Mikkeli, Finland (announced late August, site banner). Source: Investors – Cerebras And Compute Nordic Finland Announce New 165 Mw Ai Data
• Transparency publication: the model catalog now documents compression state per model (no pruned models on public endpoints; selective weight-only quantization in storage with full-precision activations and KV cache). Source: Inference-Docs – Overview

Tracking status: HLD paused Cerebras tracking as of July. August confirms the pause conditions persist: the shared catalog shrank to two models and the coding subscription has been unavailable for four consecutive months. Recommend keeping tracking paused; revisit when subscriptions reopen or a new shared coding-class model is added.

Plans

Cerebras Code Pro ($50/mo) and Max ($200/mo): still not offered. Source: Cerebras – Pricing

API Pricing

Shared-endpoint per-token rates: GPT OSS 120B and Gemma 4 31B at their published rates (unchanged; see developer-tier pricing table). Cached input tokens remain billed at the same rate as fresh input (no cache discount), with a dual-bucket rate-limit system giving 3x TPM headroom for cached tokens. Source: Cerebras – Pricing , Inference-Docs – Pricing

Model Performance / Benchmarks

No new benchmarks published. Standing speeds: ~3,000 tok/s (GPT OSS 120B), ~1,850 tok/s (Gemma 4 31B). Source: Inference-Docs – Overview

Latest News

The Finland data centre (165 MW, Mikkeli) is the month's only substantive announcement, extending capacity for the inference business. Source: Investors – Cerebras And Compute Nordic Finland Announce New 165 Mw Ai Data

Community Signals

No significant HN/Reddit threads on Cerebras found in August. Community attention to Cerebras-hosted models flows through third-party references (e.g., SWE-1.7's 1000 TPS serving claim on Cerebras hardware from July's Cognition launch).

Enterprise Readiness

Transparency Gaps