Cerebras Code now supports GLM 4.6 at 1000 tokens/sec

Original Article ↗ Hacker News Discussion ↗

Performance & Technical Claims

  • 1000 tokens/sec refers to output speed; users report code “flashing” onto the screen and workflows where waiting is more about tests/compiles than model generations.
  • Cerebras and others are said to avoid quantization; commenters attribute speed to the wafer-scale chip keeping weights and KV cache in on-chip SRAM, trading high cost per token for extreme bandwidth.
  • Some argue you can test for quantization by comparing benchmark performance across providers; others point out real evidence is limited and vendor claims aren’t easily verifiable.
  • Lack of prefix caching is suspected (or at least not visible) given the architecture, making repeated long contexts expensive.

Speed vs Quality

  • Many emphasize that raw speed transforms interaction style: more rapid refactors, UI tweaks, and “semi-interactive” workflows where an agent edits many files per call.
  • Others find GLM 4.6 “smart enough but not frontier level,” often still preferring Claude/Codex for deep reasoning, complex bugs, planning, or non-mainstream domains (embedded, UEFI, some Rust/embedded HAL tasks).
  • Multiple users say GLM 4.6 is roughly Sonnet-ish: sometimes better, sometimes worse; code can be messier and may need cleanup by a higher-quality model.

Pricing, Value, and Limits

  • $50/month (and especially $200/month) is polarizing: for some, trivial vs dev salaries and justified by preserved focus; for others, “Herman Miller” pricing for SaaS.
  • Several point out Cerebras is cheaper than some competitors on a per-token basis, but per-minute request caps and daily token ceilings are easy to hit with fast, agentic workflows.
  • Some prefer cheaper options (e.g., GLM directly via other providers) or pay-per-token, questioning what Cerebras adds beyond speed.
  • Plans and GLM 4.6 access briefly showed as “sold out,” and some users report recent queueing/lag before responses.

Workflows & Tooling

  • Popular pattern: pair a slower frontier “planner” (Claude/GPT/Gemini) with Cerebras+GLM as a fast “executor” in tools like Cline, RooCode, OpenCode, or custom TUI setups.
  • Fast models shine for: UI tweaks via voice, multi-variant component generation, quick scripting, and “AI-first” greenfield web apps.
  • Limitations noted: unstable service, no/limited search or vision in some setups, frequent retries under “high demand,” and non-trivial token burn in agentic flows.

Broader Reflections on AI Coding

  • Strong debate over “vibe coding” vs disciplined LLM-assisted development: many insist careful review, tests, and static analysis are essential, especially off the happy-path (embedded, novel domains).
  • Several commenters report previously being skeptical of AI coding, but say extremely fast, “good-enough” models finally provided a genuine productivity shift.