peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP

card: Q3_K MTP head — sizes and head claim refreshed

Q8_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q6_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q5_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q4_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q4_K_S: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

IQ4_XS: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q3_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

IQ3_XXS: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

Q2_K_XL: Sharp v22.5.0 template + Q3_K MTP head (was Q8_0)

card: CyberTiel is out — link it from the intro blockquote

assets: refresh plates to the current local build

Card: full 9-tier ladder at MTP sizes, RAM+VRAM run guidance, proofread

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

Re-embed Sharp template qwen3.8-froggeric-v22.4.0

UD-Q8_K_XL with the trained MTP head

UD-Q6_K_XL with the trained MTP head

UD-Q4_K_S with the trained MTP head

UD-Q3_K_XL with the trained MTP head

UD-Q2_K_XL with the trained MTP head

UD-IQ4_XS with the trained MTP head

UD-IQ3_XXS with the trained MTP head

Card: link the non-MTP GGUF and MLX siblings from the opening

Card: MTP section describes what we measured and how to tune, rather than prescribing settings

Model card

Vision projector, shared by both tiers

card asset: tiel_banner_eyebrow

card asset: card_tiel_swe

card asset: card_tiel_mmlu

card asset: card_tiel_local

card asset: card_tiel_claw

UD-Q5_K_XL with the trained MTP head

UD-Q4_K_XL with the trained MTP head

initial commit