# GLM Handbook

> Practical English-language guides to open models, APIs and inference.

Canonical origin: https://glmhandbook.dev

## Editorial information

- [About](https://glmhandbook.dev/about/): scope and editorial byline.
- [Methodology](https://glmhandbook.dev/methodology/): commissioning disclosure and measurement limits.
- [Sitemap](https://glmhandbook.dev/sitemap.xml)
- [RSS](https://glmhandbook.dev/feed.xml)

## Articles

- [Fast tokens. Slow answers.](https://glmhandbook.dev/articles/llm-latency-tokens-per-second/): Understand time to first token, decode speed, p95 latency and useful-task throughput before comparing LLM inference benchmarks. [Markdown](https://glmhandbook.dev/markdown/llm-latency-tokens-per-second.md)
- [Best abliterated LLM APIs](https://glmhandbook.dev/articles/best-abliterated-llm-apis/): Choose an abliterated LLM API by model access, cost and evidence. A practical guide to Uncensured AI, GPU rental and self-hosting. [Markdown](https://glmhandbook.dev/markdown/best-abliterated-llm-apis.md)
- [Make the model fit.](https://glmhandbook.dev/articles/llm-quantization-memory-guide/): Estimate LLM weight memory, understand 4-bit tradeoffs and leave room for the KV cache before choosing a local model. [Markdown](https://glmhandbook.dev/markdown/llm-quantization-memory-guide.md)
- [Abliterated LLMs, via API](https://glmhandbook.dev/articles/abliterated-llm-api-without-gpu/): Connect to abliterated GLM and DeepSeek models without expensive hardware. Uncensured AI setup, Python example, model IDs and practical checks. [Markdown](https://glmhandbook.dev/markdown/abliterated-llm-api-without-gpu.md)
- [What abliteration changes](https://glmhandbook.dev/articles/abliteration-vs-fine-tuning/): A clear explanation of abliterated LLMs, refusal directions, fine-tuning and the evaluation needed to separate compliance from capability. [Markdown](https://glmhandbook.dev/markdown/abliteration-vs-fine-tuning.md)
- [Compatible, up to a point](https://glmhandbook.dev/articles/openai-compatible-api-checklist/): Test Chat Completions compatibility, streaming, tool calls, JSON, errors and versioning before swapping an LLM endpoint into production. [Markdown](https://glmhandbook.dev/markdown/openai-compatible-api-checklist.md)
- [A big context can forget.](https://glmhandbook.dev/articles/long-context-is-not-memory/): Build better long-context LLM workflows with retrieval, evidence placement, compact state and tests that catch missed facts. [Markdown](https://glmhandbook.dev/markdown/long-context-is-not-memory.md)
- [Do the inference maths.](https://glmhandbook.dev/articles/api-vs-local-inference-cost/): Compare token-based API costs with local or rented inference using an interactive calculator and explicit workload assumptions. [Markdown](https://glmhandbook.dev/markdown/api-vs-local-inference-cost.md)
- [Test the work, not the vibe.](https://glmhandbook.dev/articles/llm-evaluation-that-catches-regressions/): Create a compact LLM evaluation suite with real tasks, explicit scoring, held-out prompts and reproducible model comparisons. [Markdown](https://glmhandbook.dev/markdown/llm-evaluation-that-catches-regressions.md)

## Evidence note

Two commissioned API guides cover Uncensured AI. The 250–300 tokens/s and 0% refusal figures were supplied by the commissioning owner and have not been independently reproduced by this publication. Read the methodology before citing them as measurements.