Qwen3.5-0.8B-ROCmFP4-GGUF

Qwen3.5-0.8B — ROCmFP4 / ROCmFPX GGUF

First ROCmFP4/ROCmFPX quantization of Qwen/Qwen3.5-0.8B. Four quants, built from the Hub BF16 GGUF (no re-conversion), each individually verified on real hardware.

Built for AMD Strix Halo (gfx1151) — Ryzen AI MAX+ 395, 128 GB unified — with the ROCmFPX llama.cpp fork.

⚠️ Needs a ROCmFPX-capable llama.cpp build. These will not load in stock llama.cpp / Ollama / LM Studio.

Variants

mmproj-BF16.gguf is included — required for image input (-fa off).

Which to pick: the 4-bit (102) — fastest and smallest, same answers as the 8-bits.

⚠️ Q6_0_ROCMFPX_AGENT (114) is the weakest choice here: larger than the 4-bit and slower than the plain 8-bit. The AGENT recipe raises speculative-draft acceptance, and Qwen3.5-0.8B ships no drafter, so that benefit cannot apply. It is included for completeness.

⚠️ Read this before deploying the 0.8B

On some prompts this model never emits a final answer — it continues reasoning until it hits the token limit. Measured on the plain 8-bit with the question "How many days were there in the year 2024?":

Eight times the budget produced eight times the reasoning and still no answer. This is not quantization damage — the BF16 source behaves identically, and the 6-bit build happens to answer the same question fine. It is the model at this size failing to terminate its own chain of thought.

Practical impact: you will see empty responses (content: "", finish_reason: "length"), not wrong ones. If you deploy this, handle empty content explicitly, and consider constraining or disabling the thinking template for short factual queries.

The correctness scores in the table above are measured with max_tokens: 1024 and count a non-terminating answer as a miss.

Measured — not estimated

Hardware: Ryzen AI MAX+ 395 (Strix Halo, gfx1151), 128 GB unified. -ngl 999 -c 4096 -fa on -fit off. 300 tokens, warm-up discarded, median of 3.

No MTP/EAGLE drafter ships with this model, so these are the deployed numbers with nothing left on the table.

Head protection

Qwen3.5-0.8B has tied embeddings — there is no output.weight tensor, so --output-tensor-type is a silent no-op and --token-embedding-type is the only flag that protects the head. Audited by exact tensor name on every artifact.

    501452832  Qwen3.5-0.8B-Q4_0_ROCMFP4_COHERENT.gguf
    734920736  Qwen3.5-0.8B-Q6_0_ROCMFPX_AGENT.gguf
    796293152  Qwen3.5-0.8B-Q8_0_ROCMFPX.gguf
    804321312  Qwen3.5-0.8B-Q8_0_ROCMFPX_AGENT.gguf

Credits

• Base model: Qwen/Qwen3.5-0.8B
• Quantization stack: ROCmFPX