license: cc-by-nc-4.0
base_model: gordonhu/Tev-320B
base_model_relation: quantized
tags:
- gguf
- project-maya
- glm-5.3-flash
- moe
- abliterated
- uncensored
- agentic
TEV-320B
⬛The quant
gordonhu/Tev-320B with dealignai's refusal edit transferred onto its BF16 weights, quantized with Project Maya's own quantizer with a custom layout (thus Claude called it nimbz-m) for my specific system (System: 2x24GB GPU, 256GB RAM, GGUF: 118.4 GB, 3 shards). Runs on Project Maya:
maya.sh --setup --gguf-dir <this folder>. I calibrated it on English and German chat, roleplay, creative writing, agentic coding + tool calls, web/general programming, and reasoning.☰Recipe
source: gordonhu/Tev-320B
abliteration: dealignai/GLM-5.3-Flash-UNCENSORED-FP8
quantizer: project-maya tools/maya_quant
calibration: 128 x 2048 tokens, rendered with Tev's stock template
recipe: nimbz-m
default: Q6_K
rules:
- { match: '^blk\.45\.ffn_(gate|up)_exps', type: Q3_K }
- { match: '^blk\.45\.ffn_down_exps', type: Q4_K }
- { match: '^blk\.([3-6]|4[1-4])\.ffn_(gate|up)_exps', type: IQ3_XXS }
- { match: 'ffn_(gate|up)_exps', type: IQ2_S }
- { match: '^blk\.([3-6]|4[1-4])\.ffn_down_exps', type: IQ3_S }
- { match: 'ffn_down_exps', type: IQ3_XXS }
- { match: 'ffn_gate_inp|exp_probs_b|hc_|indexer_compressor_ape', type: F32 }
- { match: 'ssm_(beta|f_a|f_b|g_a|g_b)|indexer', type: Q8_0 }
◖ · ◗