refusal rate
Subscribe

Open models for developers · Est. 2026

An independent guide to abliterated, open-weight and self-hosted LLMs for people who write code. How they work, how to run them on your own hardware, what each API costs, and where they still fall short.

Start reading Weekly digest
[ fig. 00 — the open-weight landscape ]44×10
[ fig. 00 — open weights ]16×9
LlamaMeta QwenAlibaba MistralMistral AI DeepSeekDeepSeek GemmaGoogle PhiMicrosoft GLMZhipu KimiMoonshot DolphinCognitive Comp. HermesNous Research OLMoAi2 GraniteIBM

Featured

[ fig. 01 — explainer ]22×11

Oct 25, 2026 · Explainer

What is abliteration, and what does an abliterated model give up?

Abliteration is a weight edit that removes a model's tendency to refuse. Research in 2024 found that refusal in chat models is largely carried by a single direction in the model's activations; projecting that direction out of the weights stops most refusals without retraining. The trade-off is a measurable drop in benchmark quality, which builders usually repair with a short preference-tuning pass.

Read the guide

Index

5 topics · 4 guides

  1. 01 Open models for coding Which open-weight and abliterated models are worth using for real coding work, and how they compare. 1 guide →
  2. 02 Abliteration & refusals How refusal behavior is represented in a model, what abliteration changes, and what it costs in quality. 1 guide →
  3. 03 Run it locally Ollama, LM Studio, llama.cpp and vLLM setups for private coding assistants on your own GPU. 2 guides →
  4. 04 Uncensored APIs Providers serving open and abliterated models over OpenAI-compatible APIs: price, privacy, limits. 0 guides →
  5. 05 Security research How security teams use open models in authorized work, and the guardrails, laws and norms that apply. 0 guides →

Latest

[ fig. 02 — best for ]18×8

What is the best local LLM for coding in 2026?

Oct 27, 2026 · Best for

For a 24GB GPU, the best local coding model in October 2026 is Qwen3.8 27B: Apache-2.0, 256K context, an 18GB Ollama download, and the strongest vendor-reported coding scores of any model that fits. Smaller tiers drop to Devstral Small 2 (16GB) and Qwen3.5 9B (8GB). Larger tiers open up mixture-of-experts models like Laguna XS 2.1 and Qwen3-Coder-Next.

Read
[ fig. 03 — how-to ]18×8

How do you run an abliterated model in Ollama or LM Studio?

Oct 27, 2026 · How-to

An abliterated model is an ordinary open-weight model with its refusal direction edited out, so it runs in Ollama and LM Studio like any other GGUF. In Ollama, one command pulls it straight from Hugging Face: ollama run hf.co/{user}/{repo}:{quant}. In LM Studio, search the Hugging Face repo or use lms get. Read the model card and license first.

Read
[ fig. 04 — hardware ]18×8

How much VRAM do you need to run a coding LLM locally?

Oct 27, 2026 · Hardware

Memory for a local coding LLM is weights plus KV cache plus about 1 GB of overhead. Weights are parameters times bits per weight divided by 8. At Q4_K_M a 9B model needs about 5.7 GB, a 27B about 16.5 GB and a 35B MoE about 22 GB, before context. Recent Qwen models cache only 64 KiB per token or less, so long context costs far less than it did.

Read

Weekly digest

New open-weight releases, evals and setup notes. One email a week.

The newsletter opens at launch. Until then, follow by RSS.

Subscribe via RSS