Applied AI Research
We research cognitive training for small local models, specialising in on-device reasoning and context optimisation for embedded systems.
SOTA LLM Inference Guide (April 2026)
Curated recommendations by VRAM tier for local inference (Q4 Quantisation).
ENTERPRISE (64GB+ VRAM)
massive scale — system-wide refactoring & agentic swarmsdeepseek-v4-pro~42 GB93.5% LCB. Best for system-wide refactoring.
qwen3.6-plus~42 GB1M Context. Optimized for agentic swarms.
llama4-maverick~40 GBMeta's flagship for multi-language & images.
LARGE (32GB VRAM)
large — gold standard logic & engineeringgemma4:31b~18 GB80% LCB; 256K context. The local logic gold standard.
qwen3.5-coder:32b~19 GBDeep specialized engineering logic.
deepseek-v4-flash~19 GB1M context; high-efficiency MoE.
MID-RANGE (16GB VRAM)
mid-range — balanced intelligence & contextgemma4-moe:26b~15 GB4B active; high-speed with 30B+ intelligence.
phi4:14b~9 GBSOTA reasoning-per-parameter for complex math/code.
mistral-small:24b~14 GBBest balanced 128k context generalist.
EFFICIENT (8GB VRAM)
efficient — real-time local intelligenceqwen3.5-coder:7b~4 GBBest real-time IDE local completions.
gemma4-e4b~6 GBNative Audio/Vision for visual debugging.
deepseek-r1:8b~5 GBBest for logical debugging and 'thinking' tasks.
LOW-POWER (4GB VRAM)
small — edge instruction followingllama4-scout:3b~2 GBSOTA sub-5GB instruction following.
qwen3.5-coder:3b~2 GBLocal 'Ghost Text' and shell scripts.
gemma3:4b~2.5 GBEfficient summarization and baseline assistant.
EDGE/MICRO (500MB - 2GB VRAM)
ultra-small — specialized agentic tasksgemma4-e2b~1.5 GBAgent-ready; native tool-use and vision.
qwen3.6:0.6b~0.5 GBSOTA sub-1GB logic for boilerplate and regex.
deepseek-r1:0.5b~0.5 GB'Thinking' minis for CLI and automation.
smollm3:500m~0.5 GBBest for data cleaning, formatting, and JSON tasks.