LFM 2.5 8B ThinkingCap is a distilled, reasoning-calibrated, and abliterated hybrid architecture combining 36 layers of 1D short-convolutions with a 32-expert MoE backbone (~1.2B active parameters per token out of 8.47B total).

Trained using full Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), this model enforces the ThinkingCap contract:

Focused internal reasoning enclosed strictly within [Start thinking] ... [End thinking].
Clean, structured, and deterministic answers in the deliverable block.
High fidelity and fluency in English and Spanish.

⚡ Quantization & APEX Matrix

This repository provides standard GGUF quantizations alongside APEX mixed-precision and native ROCmFP4 quants specifically tailored for the AMD Strix Halo (RDNA 3.5 / gfx1151 / Radeon 8050S / Radeon 8060S) platform with unified LPDDR5X memory.

Available GGUF Files

File Name	Size	Type / Description	Target Architecture
`LFM2.5-8B-A1B-ThinkingCap-ROCMFPX-APEX.gguf`	4.78 GB	Definitive ROCmFPX APEX: Routed experts in `Q4_0_ROCMFP4`, attention in `Q6_K`, shared experts/conv in `Q8_0`	AMD Strix Halo (`gfx1151`) ROCmFPX
`LFM2.5-8B-A1B-ThinkingCap-APEX.gguf`	4.63 GB	Universal APEX: Routed experts in `Q4_K_M` + `apex.txt` + `iMatrix`, attention in `Q6_K`, conv in `Q8_0`	llama.cpp / Vulkan / CPU / CUDA
`LFM2.5-8B-A1B-ThinkingCap-APEX-Q6.gguf`	6.60 GB	APEX Q6: Maximum reasoning retention with moderate compression	High-precision inference
`LFM2.5-8B-A1B-ThinkingCap-Q8_0.gguf`	8.40 GB	Full 8-bit reference	Reference evaluation
`LFM2.5-8B-A1B-ThinkingCap-Q6_K.gguf`	6.50 GB	Standard Q6_K	General purpose
`LFM2.5-8B-A1B-ThinkingCap-BF16.gguf`	16.0 GB	Unquantized BF16 Base	Research / Conversion base
`LFM2.5-8B-A1B-DSpark-Q4_K_M.gguf`	191 MB	DSpark Speculative Drafter (Q4_K_M)	Speculative decoding drafter
`LFM2.5-8B-A1B-DSpark-BF16.gguf`	633 MB	DSpark Drafter (BF16)	Speculative decoding base

📊 Measured Benchmarks (AMD Radeon 8050S / Strix Halo gfx1151)

All benchmarks measured locally on AMD Ryzen AI Max 385 APU with 64 GB LPDDR5X Unified RAM.

1. Prefill & Generation Context Ladder (`llama-bench`)

Metric	Context Size	GPU Performance
Short Prefill (`pp512`)	512 tokens	3,061.1 tokens/sec
Medium Prefill (`pp1024`)	1,024 tokens	3,165.7 tokens/sec
Long Prefill (`pp4096`)	4,096 tokens	3,564.4 tokens/sec
Deep Prefill (`pp16384`)	16,384 tokens	2,233.8 tokens/sec
Decode / Generation (Single-Stream)	128 – 4,096 tokens	115.8 – 148.3 tokens/sec

2. Multi-Slot Continuous Batching Scaling (256K Global Context)

Concurrency	Per-User Decode Speed	Aggregate System Throughput	Scaling Factor
1 Slot	148.3 tokens/sec	121.5 tokens/sec	1.00× (Baseline)
2 Slots	75.2 tokens/sec	136.4 tokens/sec	1.12×
4 Slots	64.8 tokens/sec	211.8 tokens/sec	1.74×

🎯 Recommended Sampling Parameters (DRY Sampler Sweetspot)

For optimal reasoning output and zero repetitive loops, use the DRY (Don't Repeat Yourself) sampler configuration instead of static repetition penalty:

json

{
  "temperature": 0.2,
  "top_p": 0.95,
  "dry_multiplier": 0.8,
  "dry_base": 1.75,
  "dry_allowed_length": 2,
  "max_tokens": -1
}

🚀 Quickstart & Deployment

Run with `llama-cli` (ROCm / HIP for gfx1151)

bash

export HSA_OVERRIDE_GFX_VERSION=11.5.1
export GGML_HIP_ENABLE_UNIFIED_MEMORY=1

llama-cli \
  -m LFM2.5-8B-A1B-ThinkingCap-ROCMFPX-APEX.gguf \
  -dev ROCm0 \
  -ngl 999 \
  -fa on \
  -ctk q8_0 -ctv q8_0 \
  -b 8192 -ub 2048 \
  -c 32768 \
  --dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 \
  --temp 0.2 \
  --jinja \
  -p "<|im_start|>system\nEres un asistente experto.\n<|im_end|>\n<|im_start|>user\nHola, ¿cuál es tu arquitectura?\n<|im_end|>\n<|im_start|>assistant\n"

Production Server (Docker)

bash

docker run -d \
  --name strix-halo-server-fast \
  --restart unless-stopped \
  --device=/dev/dri \
  --group-add 983 --group-add 987 \
  -p 8080:8080 \
  -v /path/to/models:/models:ro \
  strix-halo-llamacpp:ultimate-vulkan \
  llama-server \
    -m /models/LFM2.5-8B-A1B-ThinkingCap-ROCMFPX-APEX.gguf \
    -ngl 999 \
    --host 0.0.0.0 \
    --port 8080 \
    -fa 1 \
    -ctk q8_0 -ctv q8_0 \
    -b 16384 -ub 2048 \
    -np 4 -c 262144 \
    --jinja \
    --metrics

📄 License & Credits

Base Weights: Based on Liquid AI LFM 2.5 8B A1B and huihui-ai abliterated.
Fine-Tuning: ThinkingCap SFT & DPO training by osk-arr00.
ROCmFPX Kernels: Built with ROCmFP4 custom Wave32 ISA kernels for AMD RDNA 3.5 (gfx1151).

osk-arr00/LFM2.5-8B-A1B-ThinkingCap-GGUF

🧠 LFM 2.5 8B ThinkingCap — GGUF & ROCmFPX Suite

📌 Model Overview