Stock llama.cpp will not load this file. You need both the instella architecture
and the ROCmFP4 tensor types in one tree. Upstream
charlie12345/ROCmFPX has the ROCmFP4 types but
not instella. Our fork has both:
Verified 2026-08-27 on gfx1151: clean clone → 0 build errors → llama-server loads a
instella ROCmFP4 GGUF from this family and generates coherent text.
The first ROCmFP4 build of amd/Instella-MoE-16B-A3B-Think,
AMD's fully-open Mixture-of-Experts reasoning model (16B total / ~2.8B active), tuned for
AMD Strix Halo (gfx1151).
Produced end to end from AMD's original bf16 safetensors: an independent LLM_ARCH_INSTELLA
implementation written against llama.cpp upstream master 74ce157, our own GGUF conversion, and
ROCmFP4 quantization via ROCmFPX.
Why ROCmFP4 here — smaller and faster, same answers
Measured on a Ryzen AI MAX+ 395 (Strix Halo, gfx1151, 125 GiB unified), ROCm/HIP,
-ngl 999 -dio --no-warmup, 3-run medians, nonce-prefixed prompts:
Q4_K_M
ROCmFP4 STRIX
size
9.75 GiB
7.94 GiB — 19% smaller
decode
~69 tok/s
80.5 tok/s — 16.7% faster
Correctness spot-check (greedy, 5/5):17×23 → 391 · bat-and-ball → 0.05 ·
9.11 vs 9.9 → 9.9 · capital of Japan → Tokyo · Hamlet → William Shakespeare.
The two trap questions matter: both are classic failure cases where quantization damage or a
mis-implemented architecture shows up first. This build gets both right.
⛔ Requirements — read before downloading
1. You need a ROCmFPX runtime. ROCmFP4 uses ggml tensor types 100–106. Stock llama.cpp does not
implement them and will refuse the file at load:
gguf_init_from_reader: tensor 'output.weight' has invalid ggml type 101. should be in [0, 43)
Build ROCmFPX — both llama-quantizeandllama-server.
2. You need the instella architecture. It is not in llama.cpp mainline (as of master
74ce157). Instella-MoE declares model_type: deepseek_v3 and reuses DeepSeek MLA shapes, but it is
not a drop-in DeepSeek-V3:
FarSkip-Collective alters inference math — it requires dual residual streams. A naive
"DeepSeek-V3 + a gate" implementation loads without error and emits fluent-but-wrong text.
Gated MLA applies an input-conditioned gate before the output projection.
Plus rope_interleave and qk_layernorm.
3. For clean reasoning output, apply llamacpp-instella-thinkfix.patch (included).
Instella's chat template contains no <think> marker and no enable_thinking switch, so llama.cpp's
differential tag-detection finds nothing and the raw <think> block leaks into message.content.
The patch extends the legacy-DeepSeek workaround in common/chat-diff-analyzer.cpp to force a
complete <think>/</think> pair — it also fixes DeepSeek-R1, R1-Distill, and pre-DSML V3.
Then serve with --reasoning-format deepseek to get reasoning_content populated.
Environment: HSA_OVERRIDE_GFX_VERSION=11.5.1, GGML_HIP_ENABLE_UNIFIED_MEMORY=1.
Use the model's native chat template.
⚠️ Limitations (measured, not guessed)
This is a heavy reasoner — give it a large max_tokens.
It spent ~3,400 reasoning tokens on "17 × 23" and 4,615 on a two-step time question. With a
small budget you get empty content, because thinking consumes the entire allowance.
Use max_tokens ≥ 2500; prefer 4000+.
Some open-ended puzzles do not converge. A classic rope-burning puzzle consumed 6,000 tokens with
finish_reason: length and never produced an answer. Not a crash and not a wrong answer — it simply
does not terminate on certain problems.
No tool calling. The tokenizer contains no tool tokens and the chat template has no block that
injects tool definitions. Even with definitions forced into the prompt (verified at
prompt_tokens=144) the model answers in prose and never emits a tool call. Do not substitute a
DeepSeek-R1 template to try to enable it — that changes the prompt format the model was trained on
and still produces no tool calls.
Included
Instella-MoE-16B-A3B-Think-Q4_0_ROCMFP4_STRIX.gguf — the model
llamacpp-instella-thinkfix.patch — the reasoning-extraction fix described above
License
Inherits the license of the base model, amd/Instella-MoE-16B-A3B-Think.
Other public builds of this model
Compiled from Hugging Face repository metadata — file sizes, shipped files, quant variant as named by each repo. No third-party build was run or benchmarked here, so this table makes no speed or quality claim about any of them. It is here so you can see the size and format options at a glance and pick what fits your hardware.
This build would not exist without the work below. Please star and follow these
projects — the quantisation format used here is their engineering, not mine.
ROCmFPX — maintained by
charlie12345 / caf
The ROCmFP4 / ROCmFPX tensor formats (ggml types 100–106) exist only in this fork.
Every ROCmFP4 file in this repository was produced with its llama-quantize, and
runs on its runtime. The fork also credits collaborators ciru-ai, Tom Turney,
PlunderStruck and Aydan S., and acknowledges AMD for hardware support.
Licensed MIT, based on upstream llama.cpp.
llama.cpp — ggml-org and contributors
The inference engine, GGUF format and conversion tooling everything here is built on.
AMD ROCm
The compute platform these builds target — ROCm 7.2.4 on gfx1151 / Radeon 8060S.
Base model authors — see base_model in the metadata above; all model weights,
licences and capabilities are theirs. This repository contributes quantisation and
measurement only.
If you use these files, please credit ROCmFPX alongside this repository.
