lunar-V1
A local-first thinking coder with eyes — forged on a single GTX 1050 4GB through custom quantization, crowned by benchmark deathmatches, and verified task by task with everything executed for real. No cloud, no subscriptions, no telemetry: download it and it works offline.
- Author: not1Emerald
- Base: Qwen/Qwen3-VL-4B-Instruct (post-trained instruct, thinking + tools + 100+ languages)
- This file: custom Q5_K_M quant with our own Lithuanian+code importance matrix
- Context: 262,144 native architecture (serve 2–25k on 4GB VRAM)
- License: Apache-2.0 (inherited from Qwen3)
Files
lunar-V1.gguf(2.7GB) — the model. Load fully on a 4GB GPU.lunar-V1.mmproj.gguf(433MB) — vision projector (its eyes). Required for images/video.lunar-template.jinja— server chat template: lunar-V1 identity, capability affirmation, no-refusal + no-partials rules. Use with--chat-template-file(no Modelfile needed).
Run (llama.cpp + Vulkan/CUDA, no Ollama)
llama-server -m lunar-V1.gguf --mmproj lunar-V1.mmproj.gguf \
-c 8192 -ngl 99 --jinja --chat-template-file lunar-template.jinja \
--temp 0.6 --top-k 20 --top-p 0.95 --presence-penalty 1.5 --port 8080
4GB cards: -c 2048–3072 fully on-GPU. For 8–25k context add --no-kv-offload
(KV cache lives in system RAM at ~144KB/token — ceiling grows with your RAM:
~10GB free ≈ 60k+ tokens in theory, 8–16k practical on older CPUs).
Proven (all executed live, GTX 1050)
- HumanEval: 120/120 (3 seeds × 40)
- HumanEval-20: 19/20, 90/90 follow-up
- MBPP-50: 31/50 (spec-comprehension gauntlet)
- Multilingual (bigcode HumanEvalPack tools, executed): Rust-25 + JS-25 scoring now; earlier authored round: 12/12 (Rust 5/5, JS 3/3, SQL 4/4)
- Tool calling: native
tool_callsJSON - Vision: screenshots + video frames via OpenAI image_url format
- Lithuanian + English throughout
- Speed: ~16 tok/s decode fully on-GPU (small ctx) / ~8 tok/s at 25k ctx (GTX 1050 4GB)
Identity
lunar-V1 identifies as lunar-V1 (built from the Qwen3 family) via the bundled template. Raw file without the template answers as its Qwen3 base.