llama.cpp.models.ini

# [ggml-org/Qwen3.8-27B-MTP-DFlash2]    
# same with MTP - 800t/s prefill - 20 t/s generation, acceptance rate 0.5
# model = /models/huggingface/hub/models--ggml-org--Qwen3.8-27B-GGUF/snapshots/0669b98607d47046c7c2b3f801011d54a08cfccf/Qwen3.8-27B-Q4_K_M.gguf
# mmproj =  /models/huggingface/hub/models--ggml-org--Qwen3.8-27B-GGUF/snapshots/0669b98607d47046c7c2b3f801011d54a08cfccf/mmproj-Qwen3.8-27B-Q8_0.gguf
# model-draft = /models/huggingface/hub/models--incoai--Qwen3.8-27B-DFlash2-GGUF/snapshots/51962825493a48b846b40126d35c799ac4093ad0/Qwen3.8-27B-DFlash2-Q4_K_M.gguf
# spec-type = draft-dflash
# temperature = 1.0
# top-p = 0.95
# top-k = 20
# spec-draft-n-max = 7
# image-min-tokens = 3072
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

[ornith-ai/Ornith-1.5-35B]
# 2400 t/s prefill, 54-60 t/s generation without MTP, 2200 t/s prefill 60-70 t/s generation with MTP
model = /models/huggingface/hub/models--ornith-ai--Ornith-1.5-35B-A3B-GGUF/snapshots/12393612fd4f730ff5aadc23e9b8f9648aa49ceb/Ornith-1.5-35B-Q4_K_M.gguf
mmproj = /models/huggingface/hub/models--ornith-ai--Ornith-1.5-35B-A3B-GGUF/snapshots/12393612fd4f730ff5aadc23e9b8f9648aa49ceb/mmproj-Ornith-1.5-35B-BF16.gguf
jinja = true
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
spec-draft-p-min = 0.60
spec-draft-p-split = 0.10
image-min-tokens = 3072
context-shift = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[Solstice-AI/Qwopus3.8-27B-Flash-1M]
# Qwen3.8-27B-Flash = Claude Opus 3.8 distill (Jackrong/Qwopus3.8-27B-Flash), UD-Q4_K_XL 17.6GB
# hybrid qwen35: 65 layers, full_attention_interval=4 (~17 KV layers), SSM state fixed -> 1M ctx fits
# gguf metadata: context_length=262144, NO baked rope scaling -> yarn 4x for 1M
# embedded nextn/MTP layer used as drafter, multimodal via mmproj-BF16
# 700 t/s prefill, 20-26 t/s generation, draft acceptance 0.53-0.77 (mean len up to 4.8)
# ~40GB unified used -> 20GB RAM headroom left with 1M ctx
model = /models/huggingface/hub/models--Solstice-AI--Qwopus3.8-27B-Flash-GGUF-1M/snapshots/8118ba30152dd0dddd7a3236673f2d27ba9d842f/Qwopus3.8-27B-Flash-UD-Q4_K_XL-1M.gguf
# model-draft = /models/huggingface/hub/models--Solstice-AI--Qwopus3.8-27B-Flash-GGUF-1M/snapshots/8118ba30152dd0dddd7a3236673f2d27ba9d842f/speculative/Qwopus3.8-27B-DSpark-Q8_0.gguf
mmproj = /models/huggingface/hub/models--Solstice-AI--Qwopus3.8-27B-Flash-GGUF-1M/snapshots/8118ba30152dd0dddd7a3236673f2d27ba9d842f/mmproj-BF16.gguf
ctx-size = 1048576
rope-scaling = yarn
yarn-orig-ctx = 262144
spec-type = draft-mtp
# embedded nextn/MTP layer in the UD-Q4_K_XL-1M gguf; unsloth Qwen3.8-27B gets acceptance 0.74 with it.
# DSpark external drafter tested: n-max=3 works (acceptance 0.39-0.70), but n-max 5/7 (incoai dflash2
# recipe, block_size=7/n_extract=5) OOM-crashes the child at 1M ctx - no RAM headroom for spec buffers
# n-max sweep: 3=120s wall, 4=116s, 5=108s (mean len 4.5, best), 6=135s (acceptance drops 0.41-0.50)
spec-draft-n-max = 5
temperature = 1.0
top-p = 0.95
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

# [Jackrong/Qwopus3.8-27B-Flash]
# model deleted from local cache 2026-11-09 (same model available as Solstice 1M quant)
# original author quant of Qwopus3.8-27B-Flash (Claude Opus distill), reuse Solstice mmproj-BF16
# 21-23 t/s generation, 695 t/s prefill, acceptance rate = 0.6-0.7, ~10GB unified vs ~40GB for Solstice 1M
# model = /models/huggingface/hub/models--Jackrong--Qwopus3.8-27B-Flash-GGUF/snapshots/9b90cd22b19276bc9e75c0e6d1665fa48619dd5b/Qwopus3.8-27B-Flash-MTP-Q4_K_M.gguf
# mmproj = /models/huggingface/hub/models--Solstice-AI--Qwopus3.8-27B-Flash-GGUF-1M/snapshots/8118ba30152dd0dddd7a3236673f2d27ba9d842f/mmproj-BF16.gguf
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 5
# temperature = 1.0
# top-p = 0.95
# top-k = 20
# min-p = 0.0
# image-min-tokens = 3072
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

# [byteshape/Qwen3.8-27B-IQ4_XS]
# model deleted from local cache 2026-11-08
# ShapeLearn per-tensor hybrid quant, vision via mmproj, MTP head embedded in GGUF
# look https://huggingface.co/byteshape/Qwen3.8-27B-GGUF
# 21-24 t/s generation, 700 t/s prefill, acceptance rate = 0.6-0.8
# model = /models/huggingface/hub/models--byteshape--Qwen3.8-27B-GGUF/snapshots/c9cc5b2ae2520ab77d1e5d3fae395b4140bb37e8/Qwen3.8-27B-IQ4_XS-3.84bpw.gguf
# mmproj = /models/huggingface/hub/models--byteshape--Qwen3.8-27B-GGUF/snapshots/c9cc5b2ae2520ab77d1e5d3fae395b4140bb37e8/mmproj-bf16.gguf
# jinja = true
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 3
# temperature = 1.0
# top-p = 0.95
# top-k = 20
# min-p = 0.0
# image-min-tokens = 3072
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

# [byteshape/Qwen3.8-27B-IQ4_XS-DFlash2]
# model deleted from local cache 2026-11-08
# same quant with external DFlash2 draft (incoai), text-only - no mmproj (image requests fail 500)
# look https://huggingface.co/byteshape/Qwen3.8-27B-GGUF (DFlash 2 section)
# 27-32 t/s generation (+30% vs MTP), 712 t/s prefill, acceptance rate = 0.4-0.5 (blocks of 8, mean len 3.7-4.4)
# model = /models/huggingface/hub/models--byteshape--Qwen3.8-27B-GGUF/snapshots/c9cc5b2ae2520ab77d1e5d3fae395b4140bb37e8/Qwen3.8-27B-IQ4_XS-3.84bpw.gguf
# model-draft = /models/huggingface/hub/models--incoai--Qwen3.8-27B-DFlash2-GGUF/snapshots/51962825493a48b846b40126d35c799ac4093ad0/Qwen3.8-27B-DFlash2-Q4_K_M.gguf
# jinja = true
# ctx-size = 262144
# spec-type = draft-dflash
# spec-draft-n-max = 7
# temperature = 1.0
# top-p = 0.95
# top-k = 20
# min-p = 0.0
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

[prism-ml/Ternary-Bonsai-2-27B-PQ2_0]
# Ternary Bonsai 2 27B: Qwen3.8-27B backbone, ternary g128 weights 1.72 bpw + Hadamard rotation,
# 98.2% of FP16 benchmark avg at 5.9-7.2GB. card https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
# PQ2_0 = 2-bit slots 7.21GB (2.13 bpw) - faster decode on Blackwell-class (GB10) + faster prefill everywhere
# needs llama.cpp:prism-full (PrismML-Eng fork, service bonsai:8091) - stock llama.cpp rejects PQ2_0/PTQ1_0
# upstream feature request (no PR yet): https://github.com/ggml-org/llama.cpp/issues/29058
# no DSpark drafter published for Bonsai 2 yet (Bonsai 1 27B got 2.45x on GB10 with one, watch SPECULATIVE.md)
# samplers per card thinking-mode recipe (carried in GGUF metadata general.sampling.*)
# benchmark DGX Spark 2026-09-18 (prism-b10687 = image llama.cpp:prism-full, CUDA sm_121a):
#   llama-bench -ngl 99 -fa on -r 5: pp512 = 1025 t/s, tg128 = 30.5 t/s
#     (Bonsai 1 27B PQ2_0 on GB10 per Bonsai-demo: 1005 pp / 29.2 tg - Bonsai 2 slightly faster)
#   real claude-code via ccr, 42k ctx (KV q8_0): ~730 t/s prefill, ~21.5 t/s generation, fluent russian
#   cold prefill 96k ctx no cache: 686 t/s, decode drops to 16 t/s at 96k (21.5 t/s at 42k) - KV q8_0 tax
# model 6.7GB + KV 262k q8_0 ~9GB + compute ~2.5GB; does NOT fit alongside stock llama.cpp:8090
#   with 2 models loaded (~62GB) - fork auto-fit aborts when n-gpu-layers pinned, stop llama.cpp first
model = /models/huggingface/hub/models--prism-ml--Ternary-Bonsai-2-27B-gguf/snapshots/6ed5e12bf84b7a63069882c91dd9e9218647d17b/Ternary-Bonsai-2-27B-PQ2_0.gguf
mmproj = /models/huggingface/hub/models--prism-ml--Ternary-Bonsai-2-27B-gguf/snapshots/6ed5e12bf84b7a63069882c91dd9e9218647d17b/Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf
jinja = true
ctx-size = 262144
temperature = 1.0
top-p = 0.95
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 2
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Qwen3.8-27B-MTP]
# look https://unsloth.ai/docs/models/qwen3.8
# 20-25 t/s generation, 700-800 t/s prefill, acceptance rate = 0.74
model = /models/huggingface/hub/models--unsloth--Qwen3.8-27B-GGUF/snapshots/4ca720788d1e01f1bff70c033e0d0028fd02e502/Qwen3.8-27B-UD-Q4_K_XL.gguf
mmproj = /models/huggingface/hub/models--unsloth--Qwen3.8-27B-GGUF/snapshots/4ca720788d1e01f1bff70c033e0d0028fd02e502/mmproj-BF16.gguf

# ctx-size = 1048576
# rope-scaling = yarn
# yarn-orig-ctx = 262144
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temperature = 0.7
top-p = 0.80
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[bartowski/Altworld_Hemmingway-1]
# look https://huggingface.co/bartowski/Altworld_Hemmingway-1-GGUF (botp repo is a duplicate of it)
# Altworld Hemmingway-1 = Qwen3.8-27B fine-tune, MTP layers embedded in every quant, text-only repo (no mmproj)
# mmproj borrowed from base unsloth/Qwen3.8-27B-GGUF (same vision tower, like Jackrong+Solstice pairing) - works, red-square test OK
# chat template injects reasoning-effort xhigh by default, thinking can be disabled
# benchmark DGX Spark 2026-09-24: 729 t/s prefill (37k claude-code prompt, up to 944 t/s cold),
# 22.6 t/s generation (tg_3s up to 25), draft acceptance 0.81 (mean len 3.43) - same class as unsloth/Qwen3.8-27B-MTP
model = /models/huggingface/hub/models--bartowski--Altworld_Hemmingway-1-GGUF/snapshots/259605e26a47f2dcc8731ae1aa7305f8047d2c6d/Altworld_Hemmingway-1-Q4_K_M.gguf
mmproj = /models/huggingface/hub/models--unsloth--Qwen3.8-27B-GGUF/snapshots/4ca720788d1e01f1bff70c033e0d0028fd02e502/mmproj-BF16.gguf
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temperature = 0.7
top-p = 0.80
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[bartowski/HemmingwayAI_Hemmingway-1.5]
# look https://huggingface.co/bartowski/HemmingwayAI_Hemmingway-1.5-GGUF
# Hemmingway-1.5 = Hemmingway-1 (Qwen3.8-27B) fine-tune, MTP layers embedded in every quant (Q4_0 draft),
# text-only repo (no mmproj), per-tensor layouts + imatrix calibration quants
# mmproj borrowed from base unsloth/Qwen3.8-27B-GGUF (same vision tower, like Altworld Hemmingway-1 pairing)
# card: quants built with llama.cpp b11969 -> llama.cpp:spark-full rebuilt from b12032 (build 213) 2026-11-02
# benchmark DGX Spark 2026-11-02: 834 t/s cold prefill (37.9k claude-code ctx, up to 1003 t/s first chunks),
# 21-24 t/s generation, draft acceptance 0.70-0.84 (mean len 3.1-3.5) - slightly faster prefill than Altworld Hemmingway-1
model = /models/huggingface/hub/models--bartowski--HemmingwayAI_Hemmingway-1.5-GGUF/snapshots/4d91b1bc050200ccfc562c0830a78536d2dbd77b/HemmingwayAI_Hemmingway-1.5-Q4_K_M.gguf
mmproj = /models/huggingface/hub/models--unsloth--Qwen3.8-27B-GGUF/snapshots/4ca720788d1e01f1bff70c033e0d0028fd02e502/mmproj-BF16.gguf
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temperature = 0.7
top-p = 0.80
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[webmp3/Sakura-Nemotron-3.5-Bolt-74B-A9B]
# look https://huggingface.co/webmp3/Sakura-Nemotron-3.5-Bolt-74B-A9B-24GB-Uncensored-GGUF
# abliterated edit (refusal direction removed from 128 language tensors) of NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB,
# nemotron_h_moe Mamba-hybrid MoE (74B/A9B), mixed IQ2_XXS/NVFP4/Q8_0 expert quants + Q8_0-edited tensors,
# Q8_0-edited is closer to NVIDIA original than F16-edited (KLD 0.019 vs 0.154, PPL +0.8%) - pick Q8
# text-only (no mmproj), no MTP/draft head -> plain decode, thinking via enable_thinking template option
# needs llama.cpp b11681+ (nemotron_h_moe + NVFP4/IQ2_XXS mixed types), native ctx 131072
# benchmark DGX Spark 2026-11-02 (llama.cpp b12032): 1267 t/s cold prefill (40.9k claude-code ctx,
# up to 1323 t/s chunks), 28.3-28.5 t/s generation plain decode - faster than Hemmingway-1.5 with MTP
model = /models/huggingface/hub/models--webmp3--Sakura-Nemotron-3.5-Bolt-74B-A9B-24GB-Uncensored-GGUF/snapshots/251bab80314d59f978c0e3502884fcd3d51ce367/Sakura-Nemotron-3.5-Bolt-74B-A9B-24GB-Uncensored-edited-Q8_0.gguf
jinja = true
ctx-size = 131072
temperature = 0.6
top-p = 0.95
top-k = 20
min-p = 0.01
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

# [agentionai/Signal-3.8-27B]
# model deleted from local cache 2026-11-08
# token-efficiency fine-tune of Qwen3.8-27B (vision kept, thinking on by default)
# card recipe https://huggingface.co/agentionai/Signal-3.8-27B-GGUF:
#   --jinja -ngl 999 -fa on -c 65536 -ctk q8_0 -ctv q8_0
#   --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.05
#   + --spec-type draft-mtp --spec-draft-n-max 4 (94% draft acceptance on JSON vs 72% base)
# KV cache MUST stay q8_0/f16 - 4-bit value cache (-ctv q4_*) causes repetitive loops
# claims -57% answer tokens / -52% thinking tokens vs base -> less output tokens per reply
# card runs single-stream -c 65536; parallel=2 -> 32k per slot fits Claude Code system prompt,
# native ctx 262144 works too (KV q8_0 ~9GB), but 65536 per card keeps RAM headroom
# measured on DGX Spark: 730 t/s cold prefill (38k prompt in 52s), 20-27 t/s generation,
# draft acceptance 0.57-0.86 (mean len 3.3-4.5) -> same speed class as Jackrong/Qwopus MTP-Q4_K_M,
# but much more compact answers (token-efficiency tune) -> less wall time per reply
# model = /models/huggingface/hub/models--agentionai--Signal-3.8-27B-GGUF/snapshots/c6ffee404184a62e91dc8bd65883174949aa6430/AP-Q4_K_XL/Signal-3.8-27B-AP-Q4_K_XL.gguf
# mmproj = /models/huggingface/hub/models--agentionai--Signal-3.8-27B-GGUF/snapshots/c6ffee404184a62e91dc8bd65883174949aa6430/mmproj-BF16.gguf
# jinja = true
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 3
# temperature = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.05
# image-min-tokens = 3072
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

# [webhie/Qwen3.8-27B-Q4-AutoRound-Code-GGUF]
# AutoRound Q4_0 mixed quant (output.weight Q6_K), code/math/science calibration
# MTP layer embedded in GGUF, 262k native context
# 380-750 t/s prefill (depends on prompt size), 22-29 t/s generation, acceptance rate 0.68-0.70
# model deleted from local cache
# model = /models/huggingface/hub/models--webhie--Qwen3.8-27B-Q4-AutoRound-Code-GGUF/snapshots/04a41723de3622e56bb499676ebaaacaa430f345/Qwen3.8-27B-Q4_0-AutoRound-Code.gguf
# mmproj = /models/huggingface/hub/models--webhie--Qwen3.8-27B-Q4-AutoRound-Code-GGUF/snapshots/04a41723de3622e56bb499676ebaaacaa430f345/mmproj-model.gguf
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 3
# temperature = 1.0
# top-p = 0.95
# top-k = 20
# min-p = 0.0
# image-min-tokens = 3072
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

[unsloth/flint]
# Recursive Flint 39.8B MoE (2.5B active) for coding agents, 256K ctx, Apache-2.0
# card https://huggingface.co/unsloth/flint-GGUF + https://unsloth.ai/docs/models/flint
# 49 layers = 35 sliding-window (513) + 14 full-attention, 256 experts top-6 + shared expert
# arch merged upstream 2026-10-21 (llama.cpp b11820, PR #30928) - needs fresh llama.cpp:spark-full
# b11820 also dropped the 'mlock' preset option (fatal on start) - keep it out of this file
# text-only (no mmproj), no MTP/draft head -> plain decode, thinking enabled by default
# samplers per card: temp 1.0 top-p 0.95 top-k 20 min-p 0, keep compact SWA cache (no swa-full)
# KV: full-attn only ~14 layers, card recommends f16 KV; SWA layers ring-buffer -> tiny
# benchmark DGX Spark 2026-10-22: 2900 t/s prefill (35.5k claude-code ctx), 51-52 t/s
# generation with 40k ctx (73 t/s tg short ctx) - fastest GGUF here without any speculation
model = /models/huggingface/hub/models--unsloth--flint-GGUF/snapshots/20468edfc38ab43d6914cc88b08cbecb05d0fc62/flint-UD-Q4_K_XL.gguf
jinja = true
ctx-size = 262144
temperature = 1.0
top-p = 0.95
top-k = 20
min-p = 0.0
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = f16
cache-type-v = f16
kv-unified = true
threads = 20
threads-batch = 20

# [AtomicChat/Nanbeige4.3-33B-AD-Q4_K_M]
# model deleted from local cache 2026-11-09
# Nanbeige4.3-33B dense agent/reasoning model, Llama arch (52 layers, GQA 8 kv heads, rope_theta 10M), ChatML,
# thinking on default (disable: chat-template-kwargs enable_thinking=false), text-only, no MTP head -> plain decode
# card https://huggingface.co/AtomicChat/Nanbeige4.3-33B-GGUF - model is unusually quantization-sensitive:
# AD-Q4_K_M (Atomic Dynamic imatrix, top-1 92.6%, KL 0.0380) keeps way more than plain Q4_K_M (89.9%, KL 0.0693)
# native ctx 131072 -> 262144 via yarn 2x; samplers per card agentic/tool recipe (thinking/chat: temp 0.6)
# router mode: ctx-size is per-slot -> 2 slots x 262144 = ~58GB KV q8_0 + 21GB model
# benchmark DGX Spark 2026-10-25: 602-667 t/s cold prefill (40k claude-code ctx / 32k unique),
# 7.1-7.5 t/s generation at 40k ctx (10.3-10.6 t/s short ctx, plain Q4_K_M same speed - pick AD for quality),
# ccr warm run 6.0 t/s wallclock output, cold run 3.9 (thinking eats token budget), fluent russian
# - dense 33B decode class, no MTP head so no speculation possible
# model = /models/huggingface/hub/models--AtomicChat--Nanbeige4.3-33B-GGUF/snapshots/580add3bac128bc55b275e81abda491748c10a18/Nanbeige4.3-33B-AD-Q4_K_M.gguf
# jinja = true
# ctx-size = 262144
# rope-scaling = yarn
# yarn-orig-ctx = 131072
# temperature = 1.0
# top-p = 0.95
# top-k = 40
# min-p = 0.0
# reasoning-preserve = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

[unsloth/NVIDIA-Nemotron-3.5-Lightning-30B]
model = /models/huggingface/hub/models--unsloth--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF/snapshots/f2d3fe3694501008786e81e5f20360cbf715496a/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-UD-Q4_K_XL.gguf
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
spec-draft-p-min = 0.1
spec-draft-p-split = 0.1
reasoning = true
reasoning-budget = 512
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.01
context-shift = true
ctx-checkpoints = 8
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Muse-Glimmer-30B]
# reasoning = true
# reasoning-budget = 512
model = /models/huggingface/hub/models--unsloth--Muse-Glimmer-30B-GGUF/snapshots/faa5b025c584459c13febfa5c59883516710ae39/Muse-Glimmer-30B-UD-Q4_K_XL.gguf
model-draft = /models/huggingface/hub/models--unsloth--Muse-Glimmer-30B-GGUF/snapshots/faa5b025c584459c13febfa5c59883516710ae39/dflash-kquant.gguf
mmproj = /models/huggingface/hub/models--unsloth--Muse-Glimmer-30B-GGUF/snapshots/faa5b025c584459c13febfa5c59883516710ae39/mmproj-kquant.gguf
temp = 1.0
top-p = 0.95
top-k = 64
jinja = true
ctx-size = 262144
spec-type = draft-dflash
spec-draft-n-max = 3
spec-draft-ngl = 999
spec-draft-type-k = f16
spec-draft-type-v = f16
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

# [unsloth/Laguna-S-2.1]
# model deleted from local cache 2026-11-09
# 850-1100 prefill, 18-19 t/s generation
# ctx-size = 1048576
# reasoning = true
# reasoning-budget = 512
# model = /models/huggingface/hub/models--unsloth--Laguna-S-2.1-GGUF/snapshots/750f92f90cf54159c4d7a610cb7b3e74498e75c6/UD-Q4_K_XL/Laguna-S-2.1-UD-Q4_K_XL-00001-of-00003.gguf
# ctx-size = 262144
# spec-type = draft-mtp
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

# [protoLabsAI/ThinkingCap-Qwen3.6-27B-heretic-MTP]
# model deleted from local cache 2026-11-08, model removed from huggingface
# (NVFP4-MTP gguf uses ggml type 142 - stock llama.cpp rejects it)
# model = /models/huggingface/hub/models--protoLabsAI--ThinkingCap-Qwen3.6-27B-heretic-MTP-GGUF/snapshots/8f0f73ee2258a29ee123fcc3535d55390257d540/ThinkingCap-Qwen3.6-27B-heretic-NVFP4-MTP.gguf
# ctx-size = 262144
# spec-type = draft-mtp
# reasoning = true
# reasoning-budget = 512
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# threads = 20
# threads-batch = 20

[deepreinforce-ai/Ornith-1.0-35B-MTP]
# 2200-2400 t/s prefill, 30-70 t/s generation, acceptance rate 0.77
# https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF/discussions/39
# chat-template-file = /models/huggingface/hub/models--froggeric--Qwen-Fixed-Chat-Templates/snapshots/23a40b0bd4d197c31d39e3c442fd2cd6100b3971/chat_template.jinja
# samplers = "top_k;top_p;min_p;temperature;typ_p"
# swa-checkpoints = 48
# cache-reuse = 256
model = /models/huggingface/hub/models--deepreinforce-ai--Ornith-1.0-35B-GGUF/snapshots/383064f72a1ef3087b779f268d3ca117eb989aac/ornith-1.0-35b-Q4_K_M.gguf
model-draft = /models/huggingface/hub/models--a4lg--Qwen3.5-35B-A3B-MTP-ONLY-GGUF/snapshots/cd7663dc2cb47916d7f8dd1b051d4a392fc9f29d/Qwen3.5-35B-A3B-MTP-ONLY-BF16.gguf
mmproj = /models/huggingface/hub/models--juan1995-dev--Qwen3.5-35B-A3B-mmproj-BF16-GGUF/snapshots/effe29413321eaf205af0d2c767fa39ad2f1238e/mmproj-BF16.gguf
jinja = true
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
spec-draft-p-min = 0.1
spec-draft-p-split = 0.1
reasoning = true
reasoning-budget = 512
image-min-tokens = 3072
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
context-shift = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Gemma4-31B-MTP]
# 15-16 t/s generation, 470 t/s prefill, accepance rate 0.5
model = /models/huggingface/hub/models--unsloth--gemma-4-31B-it-qat-GGUF/snapshots/43cc1aeb31adf47ec06a854507ce552cd9862e6f/gemma-4-31B-it-qat-UD-Q4_K_XL.gguf
model-draft = /models/huggingface/hub/models--unsloth--gemma-4-31B-it-qat-GGUF/snapshots/43cc1aeb31adf47ec06a854507ce552cd9862e6f/mtp-gemma-4-31B-it.gguf
mmproj = /models/huggingface/hub/models--unsloth--gemma-4-31B-it-qat-GGUF/snapshots/43cc1aeb31adf47ec06a854507ce552cd9862e6f/mmproj-BF16.gguf
reasoning = true
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 2
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[mradermacher/Gemma-4-Novelist-31B]
model = /models/huggingface/hub/models--mradermacher--Gemma-4-Novelist-31B-i1-GGUF/snapshots/ead8ac030330dc43317277f9145808fb16304257/Gemma-4-Novelist-31B.i1-Q4_K_M.gguf
ctx-size = 262144
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 2
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Gemma4-26B-MTP]
# 44-55 t/s generation, 1400-1800 t/s prefill
model = /models/huggingface/hub/models--unsloth--gemma-4-26B-A4B-it-qat-GGUF/snapshots/7b92b5b28818151e8669af2e45e88d6086f490dd/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf
model-draft = /models/huggingface/hub/models--unsloth--gemma-4-26B-A4B-it-qat-GGUF/snapshots/7b92b5b28818151e8669af2e45e88d6086f490dd/mtp-gemma-4-26B-A4B-it.gguf
mmproj = /models/huggingface/hub/models--unsloth--gemma-4-26B-A4B-it-qat-GGUF/snapshots/7b92b5b28818151e8669af2e45e88d6086f490dd/mmproj-BF16.gguf
reasoning = on
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temp = 1.0
top-p = 0.95
top-k = 64
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 2
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[DavidAU/Qwen3-48B-A4B-Savant-Commander-Distill-12X]
# card https://huggingface.co/DavidAU/Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored-GGUF
# Qwen3 MoE 48B (A4B active), 12X expert-expanded "Savant-Commander" distill +
# Closed-Open-Heretic uncensoring, Q4_K_M 20.4GB, no MTP head -> plain decode
# MANDATORY per discussion #2 (model card is WRONG): GGUF metadata lies about
# experts per token (gguf default = 2) -> override to 6, without it llama.cpp loads wrong expert count
# https://huggingface.co/DavidAU/Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored-GGUF/discussions/2
# this is 12x4B "MOE Dense": each expert = 4B distill of a frontier model (GPT-5.1/Gemini-3-Pro/
# Claude-4.5-Opus/Kimi-K2/gpt-oss-120b/GLM-Flash-4.6/...), expert_used_count = how many distill
# "voices" mix per token -> quality/speed dial, NO MTP possible (custom merge, no draft head exists)
# expert sweep 2026-11-09 @37.8k claude-code ctx (prefill t/s / decode t/s / ccr wallclock t/s):
#   2 experts: 1762 / 12.2 / 8.1   4 experts: 1212 / 10.1 / 5.9   6 experts: 880 / 8.8 / 5.0
model = /models/huggingface/hub/models--DavidAU--Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored-GGUF/snapshots/4cc51028f92bff2f8b743275eb7211af200035cb/Qwen3-48B-A4B-Savant-Commander-Dstll-12X-Cl-Op-Hrtic-Uncen-Q4_K_M.gguf
override-kv = qwen3moe.expert_used_count=int:6
jinja = true
ctx-size = 262144
temperature = 0.7
top-p = 0.80
top-k = 20
min-p = 0.0
image-min-tokens = 3072
reasoning-preserve = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[empero-ai/Qwen3.8-35B-A3B-Distill]
# look https://huggingface.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF
# Qwen3.8 frontier distill into Qwen3.6-35B-A3B MoE arch (30 Gated DeltaNet + 10 full-attn, 256 experts top-8),
# reasoning model (every answer opens with <think>), vision tower inherited unchanged from Qwen3.6 base (mmproj-F16)
# full MTP block embedded (block_count=41 = 40 trunk + blk.40 MTP with own attn+MoE + nextn.* tensors,
# same layout as unsloth/Qwen3.6-35B-A3B-MTP-GGUF) -> draft-mtp without external draft model
# samplers per card: temp 0.6 top-p 0.95 top-k 20
# benchmark DGX Spark 2026-11-08: 2884 t/s prefill (37.8k claude-code ctx), 65.9-69.4 t/s generation,
# draft acceptance 0.55 (mean len 2.65) at n-max 3, warm ccr run 61.7 t/s output;
# n-max 5 swept: no speed gain (58.6-65.3 t/s), acceptance drops to 0.38-0.45 -> keep 3
# re-verified on llama.cpp b12202 (0.9.0-dev build 387, image rebuilt 2026-11-08): same numbers
# (2625 t/s prefill, 64-66.5 t/s generation, acceptance 0.53-0.56, warm ccr 63.2 t/s) - no regressions
model = /models/huggingface/hub/models--empero-ai--Qwen3.8-35B-A3B-Distill-GGUF/snapshots/b1f9d1dcc3de8aa867669b0ab919384aeeb9b8d5/Qwen3.8-35B-A3B-Q4_K_M.gguf
mmproj = /models/huggingface/hub/models--empero-ai--Qwen3.8-35B-A3B-Distill-GGUF/snapshots/b1f9d1dcc3de8aa867669b0ab919384aeeb9b8d5/mmproj-Qwen3.8-35B-A3B-F16.gguf
jinja = true
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temperature = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
image-min-tokens = 3072
reasoning-preserve = true
context-shift = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Qwen3.6-35B-MTP]
# 70-90 t/s generation 2200-2400 t/s prefill, acceptance 0.8-0.9
model = /models/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-MTP-GGUF/snapshots/5bc3e238d916f48a861bac2f8a1990a0e9b7e98d/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
mmproj = /models/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-MTP-GGUF/snapshots/5bc3e238d916f48a861bac2f8a1990a0e9b7e98d/mmproj-BF16.gguf
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
context-shift = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 4
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

[unsloth/Qwen3.6-27B-MTP]
# 15-20 t/s generation - 500-600 t/s prefill
model = /models/huggingface/hub/models--unsloth--Qwen3.6-27B-MTP-GGUF/snapshots/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace/Qwen3.6-27B-UD-Q4_K_XL.gguf
mmproj = /models/huggingface/hub/models--unsloth--Qwen3.6-27B-MTP-GGUF/snapshots/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace/mmproj-BF16.gguf
ctx-size = 262144
spec-type = draft-mtp
spec-draft-n-max = 3
temp = 0.6
top-p = 0.95
top-k = 20
min-p = 0.00
context-shift = true
ctx-checkpoints = 8
n-gpu-layers = 99
batch-size = 8192
ubatch-size = 2048
parallel = 2
flash-attn = true
cache-type-k = q8_0
cache-type-v = q8_0
kv-unified = true
threads = 20
threads-batch = 20

# [Tariel/Ornith-1.5-35B-A3B-MTP]
# doesn't work
# look details in https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/discussions/10  
# model = /models/modelscope/models/Tariel--Ornith-1.5-35B-A3B-MTP-GGUF/snapshots/master/Ornith-1.5-35B-A3B-APEX-MTP-I-Quality.gguf
# model-draft =  /models/modelscope/models/Tariel--Ornith-1.5-35B-A3B-MTP-GGUF/snapshots/master/mtp-head/Ornith-1.5-35B-A3B-MTP-Head-Q4_0.gguf
# mmproj = /models/modelscope/models/Tariel--Ornith-1.5-35B-A3B-MTP-GGUF/snapshots/master/mmproj-APEX.gguf
# jinja = true
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-p-min = 0.60
# spec-draft-p-split = 0.10
# spec-draft-n-min = 0
# spec-draft-n-max = 3
# image-min-tokens = 3072
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 4
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20


# [unsloth/Gemma4-12B-MTP]
# 44-50 t/s generation, 1400-1800 t/s prefill, looks stupid, acceptance 0.5, 23-25 t/s without MTP
# model = /models/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/7102bdea62863acff919c945405ef29973113d66/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf
# model-draft = /models/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/7102bdea62863acff919c945405ef29973113d66/MTP/gemma-4-12B-it-Q8_0-MTP.gguf
# model-draft = /models/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/7102bdea62863acff919c945405ef29973113d66/mtp-gemma-4-12B-it.gguf
# mmproj = /models/huggingface/hub/models--unsloth--gemma-4-12B-it-qat-GGUF/snapshots/7102bdea62863acff919c945405ef29973113d66/mmproj-BF16.gguf
# reasoning = on
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 4
# temp = 1.0
# top-p = 0.95
# top-k = 64
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [MaralGPT/MaralGPT-Mythos-9B]
# 20 t/s, too slow
# model = /models/huggingface/hub/models--MaralGPT--MaralGPT-Mythos-9B-2606-GGUF/snapshots/94fdb18460a1279046d023d9746755fa90678760/MaralGPT-Mythos-9B-2606-Q8_0.gguf
# gpu-layers = all
# gpu-layers-draft = all
# ctx-size = 262144
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# repeat-penalty = 1.05
# n-predict = 16384
# n-gpu-layers = all
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# too creative, lost context, too stopid

# [Guilherme34/Firefly-v4]
# model = /models/huggingface/hub/models--Guilherme34--Firefly-v4-BetterGGUF/snapshots/4ec603afafc7c9433e520bb86c1c03072de60537/Firefly-v4.Q8_0.gguf
# model-draft = /models/huggingface/hub/models--Guilherme34--Firefly-v4-BetterGGUF/snapshots/4ec603afafc7c9433e520bb86c1c03072de60537/mtp-gemma-4-E2B-it.gguf
# mmproj = /models/huggingface/hub/models--Guilherme34--Firefly-v4-BetterGGUF/snapshots/4ec603afafc7c9433e520bb86c1c03072de60537/mmproj-BF16.gguf
# reasoning = on
# ctx-size = 262144
# spec-type = draft-mtp
# spec-draft-n-max = 4
# temp = 0.6
# top-p = 0.95
# top-k = 64
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# cache-type-k-draft = q8_0
# cache-type-v-draft = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [unsloth/Qwen3.6-35B]
# 44 t/s generation - 2400 t/s context parsing
# model = /models/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-GGUF/snapshots/a483e9e6cbd595906af30beda3187c2663a1118c/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
# mmproj = /models/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-GGUF/snapshots/a483e9e6cbd595906af30beda3187c2663a1118c/mmproj-BF16.gguf
# ctx-size = 262144
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [catlilface/gemma-4-26b]
# gemma
# model = /models/huggingface/hub/models--catlilface--Gemma-4-26B-A4B-NVFP4-GGUF/snapshots/da40399d7f0b4b987e6b93385b16e8ad102d1385/Gemma4-26b-NVFP4Q8.gguf
# mmproj = /models/huggingface/hub/models--catlilface--Gemma-4-26B-A4B-NVFP4-GGUF/snapshots/da40399d7f0b4b987e6b93385b16e8ad102d1385/mmproj-Gemma4-26b-F16.gguf
# ctx-size = 262144
# temperature = 1.0
# top-p = 0.95
# top-k = 64
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [unsloth/gemma-4-31B]
# 7 t/s generation - 350-450 t/s context parsing
# model = /models/huggingface/hub/models--unsloth--gemma-4-31B-it-GGUF/snapshots/965dd40a1c73d00a40f85a53bb540c1719d208ec/gemma-4-31B-it-UD-Q4_K_XL.gguf
# ctx-size = 262144
# temperature = 1.0
# top-p = 0.95
# top-k = 64
# batch-size = 8192
# ubatch-size = 2048
# parallel = 1
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [mudler/Qwen3.6-35B-Mini]
# 34 t/s generation, 820 t/s prefill, wtf
# model =  /models/huggingface/hub/models--mudler--Qwen3.6-35B-A3B-APEX-GGUF/snapshots/42c47e7a396813c593fb12c9307ada5cd8090d4b/Qwen3.6-35B-A3B-APEX-I-Mini.gguf
# mmproj = /models/huggingface/hub/models--mudler--Qwen3.6-35B-A3B-APEX-GGUF/snapshots/42c47e7a396813c593fb12c9307ada5cd8090d4b/mmproj.gguf
# ctx-size = 262144
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20

# [mudler/Qwen3.6-35B-Quality]
# model =  /models/huggingface/hub/models--mudler--Qwen3.6-35B-A3B-APEX-GGUF/snapshots/42c47e7a396813c593fb12c9307ada5cd8090d4b/Qwen3.6-35B-A3B-APEX-I-Quality.gguf
# mmproj = /models/huggingface/hub/models--mudler--Qwen3.6-35B-A3B-APEX-GGUF/snapshots/42c47e7a396813c593fb12c9307ada5cd8090d4b/mmproj.gguf
# ctx-size = 262144
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [CISCai/gemma-4-31B]
# 2 t/s generation 50 t/s context parsing? WTF?
# model = /models/huggingface/hub/models--CISCai--gemma-4-31B-it-NVFP4-turbo-GGUF/snapshots/c7ffdda99dc7e0bf4197f0180f7f54700c3db726/gemma-4-31B-it-NVFP4-turbo-NVFP4.gguf
# mmproj = /models/huggingface/hub/models--CISCai--gemma-4-31B-it-NVFP4-turbo-GGUF/snapshots/c7ffdda99dc7e0bf4197f0180f7f54700c3db726/mmproj-gemma-4-31B-it-NVFP4-turbo-bf16.gguf
# ctx-size = 262144
# temperature = 1.0
# top-p = 0.95
# top-k = 64
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [Exil01/Qwen3.5-35B-1M]
# model = /models/huggingface/hub/models--Exil01--Qwen3.5-35B-Uncensored-YaRN-1M/snapshots/905baef0417ac69c6cd67617ac777c4ab5443d94/Qwen3.5-35B-Uncensored-YaRN-1M-Q4_K_M.gguf
# mmproj = /models/huggingface/hub/models--Exil01--Qwen3.5-35B-Uncensored-YaRN-1M/snapshots/905baef0417ac69c6cd67617ac777c4ab5443d94/mmproj-uncensored-35b-f16.gguf
# ctx-size = 1048576
# rope-scaling = yarn
# yarn-orig-ctx = 262144
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 1
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [mradermacher/Nemotron-Cascade-2-30B]
# 60-70t/s generation - 2900t/s context parsing
# model = /models/huggingface/hub/models--mradermacher--Nemotron-Cascade-2-30B-A3B-i1-GGUF/snapshots/b685793198b3096b09c7d74ef9797e431c5be9f8/Nemotron-Cascade-2-30B-A3B.i1-Q4_K_M.gguf
# ctx-size = 524288
# n-predict = 32768
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 2048
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# temp = 1.0
# top-k = 40
# op-p = 0.95
# min-p = 0.05
# presence-penalty = 0.0
# [unsloth/Qwen3.5-35B]
# 49-51t/s generation, 1800-2500t/s context parsing
# look https://unsloth.ai/docs/models/qwen3.5
# model = /models/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# ctx-size = 524288
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 8192
# parallel = 2
# flash-attn = true
# swa-full = false
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [noctrex/Qwen3.5-35B]
# 39-50 t/s generation, 1500-2000 t/s context parsing
# model = /models/huggingface/hub/models--noctrex--Qwen3.5-35B-A3B-MXFP4_MOE-GGUF/snapshots/f47863e4882e413c572e071d43dfc4d3962dae58/Qwen3.5-35B-A3B-MXFP4_MOE_BF16.gguf
# multi modality encoder
# mmproj = /models/huggingface/hub/models--noctrex--Qwen3.5-35B-A3B-MXFP4_MOE-GGUF/snapshots/f47863e4882e413c572e071d43dfc4d3962dae58/mmproj-BF16.gguf
# ctx-size = 524288
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 8192
# parallel = 2
# flash-attn = true
# swa-full = false
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# temp = 0.6
# top-k = 20
# top-p = 0.95
# min-p = 0.0
# presence-penalty = 1.5
# [unsloth/Qwen3.5-27B]
# 10t/s generation, 500t/s context parsing
# look https://unsloth.ai/docs/models/qwen3.5
# model = /models/huggingface/hub/models--unsloth--Qwen3.5-27B-GGUF/snapshots/3221f178a6b842d04f1fb42f1c413534adcc0a6a/Qwen3.5-27B-UD-Q4_K_XL.gguf
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# ctx-size = 524288
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 4096
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [noctrex/Qwen3-Next-80B]
# model = /models/huggingface/hub/models--noctrex--Qwen3-Next-80B-A3B-Instruct-1M-MXFP4_MOE-GGUF/snapshots/4c791a92568d68271a1f38403b7932a518631787/Qwen3-Next-80B-A3B-Instruct-1M-MXFP4_MOE-00001-of-00003.gguf
# ctx-size = 1048576
# rope-scaling = yarn
# rope-scale = 4
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 4096
# [unsloth/GLM-4.7-Flash-30B]
# model = /models/huggingface/hub/models--unsloth--GLM-4.7-Flash-GGUF/snapshots/66205ea11ef638d83850389bd921e6c199f7504e/GLM-4.7-Flash-Q8_0.gguf
# temp = 0.7
# top-p = 1.0
# min-p = 0.01
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 4096
# outdated
# [Biomanticus/RYS-Qwen3.5-27B]
# 10 t/s generation - 400-500 t/s context parsing
# model = /models/huggingface/hub/models--Biomanticus--RYS-Qwen3.5-27B-gguf/snapshots/eb0e14f2d3c717f7692593ff3fddced66d56dba7/RYS-Qwen3.5-27B-Q4_K_M.gguf
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# ctx-size = 524288
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# batch-size = 8192
# ubatch-size = 4096
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [unsloth/Qwen3.5-122B]
# model = /models/huggingface/hub/models--unsloth--Qwen3.5-122B-A10B-GGUF/snapshots/d503adb1b5f3b29521713c7460735ed753b04e68/MXFP4_MOE/Qwen3.5-122B-A10B-MXFP4_MOE-00001-of-00003.gguf
# temp = 0.6
# top-p = 0.95
# top-k = 20
# min-p = 0.00
# ctx-size = 524288
# context-shift = true
# ctx-checkpoints = 8
# n-gpu-layers = 99
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 4096
# parallel = 2
# flash-attn = true
# cache-type-k = q8_0
# cache-type-v = q8_0
# kv-unified = true
# mlock = false
# threads = 20
# threads-batch = 20
# [Qwen/Qwen3-Coder-Next-80B-Q8]
# model = /models/huggingface/hub/models--Qwen--Qwen3-Coder-Next-GGUF/snapshots/b82fb7382639d97b38fa7672e526c760c2fb358e/Qwen3-Coder-Next-Q8_0/Qwen3-Coder-Next-Q8_0-00001-of-00004.gguf
# ctx-size = 262144
# cont-batching = true
# temp = 1.0
# top-p = 0.95
# top-k = 40
# cache-ram = 16384
# batch-size = 8192
# ubatch-size = 8192
# [rushyrush/MiniMax-M2.1-139B]
# model = /models/huggingface/hub/models--rushyrush-MiniMax-M2.1-REAP-139B-A10B-GGUF/snapshots/
# ctx-size = 167936
# don't feet im memory and bad russian
# [noctrex/MiniMax-M2.5-139B]
# model = /models/huggingface/hub/models--noctrex--MiniMax-M2.5-REAP-139B-A10B-MXFP4_MOE-GGUF/snapshots/6427bb42d0d3a4ee87070634815972cd99d281fd/MiniMax-M2.5-REAP-139B-A10B-MXFP4_MOE-00001-of-00004.gguf
# ctx-size = 196608
# ctx-checkpoints=16
# cache-reuse=2048
# batch-size=4096
# ubatch-size=1024
# parallel=2
# swa-full=true
# cont-batching=true
# stupid
# [cturan/IQuest-Coder-V1-40B]
# model = /models/huggingface/hub/models--cturan--IQuest-Coder-V1-40B-Instruct-GGUF/snapshots/c520c173b24f57d25e6431598cd2b3d2c4c8748e/IQuest-Coder-V1-40B-Instruct.Q4_K_M.gguf
# stupid
# [tiiuae/Falcon-H1R-7B]
# model = /models/huggingface/hub/models--tiiuae--Falcon-H1R-7B-GGUF/snapshots/21539e5e8334bfd5474da7fffd0f82d0a97c6d0e/Falcon-H1R-7B-Q8_0.gguf
# too stupid
# [noctrex/HyperNova-60B]
# model = /models/huggingface/hub/models--noctrex--HyperNova-60B-MXFP4_MOE-GGUF/snapshots/81f93ea3f69d38c89c8e032766573b5962185a73/HyperNova-60B-MXFP4_MOE.gguf
# too old
# [noctrex/Nemotron-3-Nano-30B]
# model = /models/huggingface/hub/models--noctrex--Nemotron-3-Nano-30B-A3B-MXFP4_MOE-GGUF/snapshots/175835a96514b1c3c12a5f4b842c829d26dba202/NVIDIA-Nemotron-3-Nano-30B-A3B-MXFP4_MOE.gguf
# too slow, but works
# [unsloth/Devstral-2-123B-Instruct-2512-GGUF]
# model = /models/huggingface/hub/models--unsloth--Devstral-2-123B-Instruct-2512-GGUF/snapshots/1f2bfbe35f7f9071d9b318374bf5eeffefab4459/UD-Q4_K_XL/Devstral-2-123B-Instruct-2512-UD-Q4_K_XL-00001-of-00002.gguf
# [bartowski/zai-org_GLM-4.7]
# model = /models/huggingface/hub/models--bartowski--zai-org_GLM-4.7-GGUF/snapshots/c47ba4573051e4a5146879669a8a95743b5302a4/zai-org_GLM-4.7-IQ1_M/zai-org_GLM-4.7-IQ1_M-00001-of-00003.gguf
# [endyjasmi/Qwen3-Embedding-8B]
# model = /models/huggingface/hub/models--endyjasmi--Qwen3-Embedding-8B-Q4_K_M-GGUF/snapshots/b74bf2a75d37d23867612c1b81655e4496535086/qwen3-embedding-8b-q4_k_m.gguf