BTL-4 Compact (35B-A3B, IQ2_XXS)
Bad Theory Labs' agentic build of Ornith 1.0 35B, tuned for tool-heavy coding and long stretches of unattended work. Native reasoning, native tool calling, and a 256K context window that stays light on memory even when it's full. Optimized for low-memory machines — the whole model fits in under 10GB, so it runs on a 16GB laptop, trading some accuracy against the full-size BTL-4 build.
At a glance
- Parameters
- 35B
- Quantization
- IQ2_XXS
- Context window
- 262K tokens
- Approx. size
- 10.0 GB
- Engines
- llama.cpp
- License
- Apache-2.0 (open)
- Version
- v1.0.2 , released 2026-08-15
- Family
- qwen
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
0.6- sampling.topP
0.95- sampling.topK
20- sampling.minP
0- sampling.maxTokens
16384- reasoning.thinkingBudget
8192- reasoning.enableThinking
true- profiles
- thinking-general, thinking-coding, thinking-precise, instruct, creative
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-condensed
- fabrication.detect-past-tense-no-tools
- turn.ollama-num-predict-bumped
- turn.preamble-folding
- mcp.compact-tool-schemas
Sources
- llama.cpp (GGUF)
- badtheorylabs/BTL-4-Compact ·
BTL-4-IQ2_XXS.gguf· IQ2_XXS - Upstream
- https://huggingface.co/badtheorylabs/BTL-4-Compact
- Manifest
- View on GitHub
Tags
- badtheorylabs
- agentic
- tools
- code
- reasoning
- long-context
- moe