⚠️ STOCK llama.cpp WILL NOT LOAD THESE

The ROCmFP4 / ROCmFPX quant types exist only in charlie12345/ROCmFPX — stock llama.cpp reports invalid ggml type. Ignore the auto-generated "Use this model" commands above.

LFM2-8B-A1B — ROCmFP4 / ROCmFPX GGUF

The first ROCmFP4 quantizations of LFM2-8B-A1B. Built for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo). 24 layers · hidden 2048 · 32 experts, ~1B active.

Quantized from BF16 GGUF — a lossless source, not a requantization of a lower-bit build.

All four quants — pick by size, not by hoping

Measured on a Ryzen AI MAX+ 395 (gfx1151, ROCm), -ngl 999 -c 4096 -fa on -fit off, median of 3 with warm-up discarded on an otherwise-idle box.

Correctness: 3/3 on every quant — 17x23 => 391 - capital of Japan => Tokyo - days in 2024 => 366 (max_tokens 1024)

Speculative decoding: None. No MTP / EAGLE / nextn tensors, so there is no speculative speedup to enable and no acceptance length to quote. Source: unsloth/LFM2-8B-A1B-GGUF BF16.

ℹ️ Why no "4-bit AGENT"? The AGENT recipes exist only at 3-bit (113), 6-bit (114/117) and 8-bit (115) — the entire 4-bit family (100–106) has no AGENT variant. Rather than hand-roll one and publish it under a name implying it is standard, the 6-bit Q6_0_ROCMFPX_AGENT fills that slot. AGENT routing keeps more tensors at higher precision for tool-call coherence.

ℹ️ Head protection. These GGUFs contain no separate output.weight — the embedding is tied — so --token-embedding-type is what protects the head here and --output-tensor-type is a no-op. Audited in the finished files: token_embd is Q6_K on the 4-bit build and Q8_0 on the 6/8-bit builds.

---

What was NOT measured

• No perplexity run, and no quality A/B against the BF16 source.
• No long-context testing. · No tool-calling evaluation — note the AGENT builds are named for tool-call routing but we did not benchmark tool use.

Base model licence inherited; all credit for the weights goes to Liquid AI.