Instructions to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
# Run inference directly in the terminal:
llama cli -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
# Run inference directly in the terminal:
llama cli -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
# Run inference directly in the terminal:
./llama-cli -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
# Run inference directly in the terminal:
./build/bin/llama-cli -hf kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Use Docker

docker model run hf.co/kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

LM Studio
Jan

vLLM

How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Ollama
How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with Ollama:
```
ollama run hf.co/kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
```

Unsloth Studio

How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF to start chatting

Docker Model Runner
How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with Docker Model Runner:
```
docker model run hf.co/kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP
```

Lemonade

How to use kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF:Q4_0_ROCMFP

Run and chat with the model

lemonade run user.Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF-Q4_0_ROCMFP

List all available models

lemonade list

Atomic Chat

Instella-MoE-16B-A3B-Think — ROCmFP4 STRIX GGUF — AMD Ryzen AI Max+ 395 / Strix Halo / gfx1151

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.

	value
file	`Instella-MoE-16B-A3B-Think-Q4_0_ROCMFP4_STRIX.gguf`
size	8,520,936,448 bytes (7.94 GiB)
sha256	`c37aaf551628a96438a455f65d2d24652779fd3249e5d4e80d45e92bb8c4e6b2`
real BPW	4.29 (advertised ~4.49 — measured, not assumed)
quant type	`105` `Q4_0_ROCMFP4_STRIX` — Strix Halo attn-K/V quality recipe

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-quantize and llama-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.

Usage

llama-server -m Instella-MoE-16B-A3B-Think-Q4_0_ROCMFP4_STRIX.gguf \
  -ngl 999 -dio --no-warmup --jinja \
  --reasoning-format deepseek \
  -c 32768 --parallel 1

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
imatrix.dat — importance matrix (37.8 MB), 98.44% expert-tensor coverage
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.

Repository	Largest model file	Variant	Ships	Downloads	Likes
`kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF` (this repo)	7.94 GiB	STRIX	single model file	12	0
`kingjones777/Instella-ToolCall-16B-A3B-ROCmFP4-STRIX-GGUF`	7.94 GiB	STRIX	safetensors	0	0

Base model: amd/Instella-MoE-16B-A3B-Think. Generated from Hub metadata; download counts move over time.

Acknowledgements

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.

Downloads last month: 138

GGUF

Model size

16B params

Architecture

instella

Hardware compatibility

4-bit

Model tree for kingjones777/Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF

Base model

amd/Instella-MoE-16B-A3B-Think

Quantized

(5)

this model

kingjones777
/

Instella-MoE-16B-A3B-Think-ROCmFP4-STRIX-GGUF