How to use from
vLLM
Install from pip and serve model
# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Harsha001/Inkling-2-Mini-ROCmFPX"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Harsha001/Inkling-2-Mini-ROCmFPX",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'
Use Docker
docker model run hf.co/Harsha001/Inkling-2-Mini-ROCmFPX:Q4_0_ROCMFP
Quick Links

Inkling 2 Mini ROCmFPX (AMD GPUs and Strix Halo)

Hardware-aware GGUF quants of Inkling 2 Mini by Thinking Machines, built with the ROCmFPX weight formats for AMD GPUs (Vulkan and ROCm).

Inkling 2 Mini: 26.1B parameters, 24 layers, 128 experts per layer with 8 active per token (about 3.1B active parameters per layer pass). It is a looped model: layers 6 to 17 run three times per token, so a token passes through 48 layers. Text only here: the vision encoder and the multi-token prediction drafter are not included.

Not an official release. All credit for the model goes to Thinking Machines.

Files

File Size Weights Fits
Inkling-2-Mini-Q8_0_ROCMFPX.gguf 26.9 GB everything 8-bit ROCmFPX (token embedding Q8_0) best quality: Strix Halo, 32 GB AMD GPUs
Inkling-2-Mini-Q4_0_ROCMFP4_STRIX_O8.gguf 15.5 GB experts ROCmFP4, everything else 8-bit ROCmFPX best smaller: 24 GB AMD GPUs, Strix Halo with room for long context

Quality

KL divergence against the original weights (Thinking Machines' NVFP4 experts kept as they are, the MXFP8 layers 20 to 23 stored as BF16), 40 chunks of 512 tokens of English Wikipedia text (wikitext-2), CPU. Lower is better. "Same top token" is how often the file's most likely next token matches the original.

Build Size Perplexity KL to original Same top token
Original weights (reference, not published) 22.9 GB 6.743
Q8_0_ROCMFPX 26.9 GB 6.741 0.0039 97.2%
Q4_0_ROCMFP4_STRIX_O8 15.5 GB 6.784 0.0120 95.4%
stock Q8_0 (comparison, not uploaded) 27.7 GB 6.739 0.0035 97.5%
stock Q4_K_M (comparison, not uploaded) 17.4 GB 6.819 0.0174 94.1%

At 4 bits the ROCmFP4 file is smaller than stock Q4_K_M and closer to the original on every measure. At 8 bits stock Q8_0 is a little closer to the original (0.0035 against 0.0039) and 0.8 GB larger; the ROCmFPX file is the one built for the AMD kernels.

The Q8 file is larger than the original checkpoint because the original experts are 4-bit NVFP4. 8 bits keep them almost exactly; the 4-bit ROCmFP4 file is the size to pick when memory is tight.

Every file also passed a greedy chat check before upload ("What is the capital of France? Answer in one word." gives "Paris" after a short thinking part).

Running it

Inkling 2 Mini needs the looped inkling2 graph, which is only in the ROCmFPX fork of llama.cpp (branch inkling2-mini). Build it with Vulkan or ROCm/HIP, then for example:

llama-server -m Inkling-2-Mini-Q4_0_ROCMFP4_STRIX_O8.gguf -ngl 99 -c 16384 --jinja

Known limits of this first version:

  • Only the CPU path has been checked against the original PyTorch model (it matches to about 5e-6 on a small test model, and the tokenizer matches exactly). Vulkan and ROCm have not been tried yet. Please report in the Community tab whether it runs, how fast, and how the answers look.
  • llama.cpp has no parser for Inkling's message format yet. Answers come back with Inkling's own markers (<|content_thinking|>, <|end_message|> and so on): the thinking is not split from the answer and tool calls are not parsed. Plain chat works; agent use does not yet.
  • The chat template accepts reasoning_effort as none, minimal, low, medium, high, xhigh or max, or as a number from 0 to 0.99 (default 0.9).

How it was made

  • Converted from the original checkpoint with the NVFP4 experts kept as they are, then requantized with llama-quantize. The per-expert NVFP4 scale is folded into the weights before requantizing.
  • ROCmFPX tensor scales on (a power-of-two scale per tensor, stored as .scale), token embedding Q8_0.
  • Importance matrix from a general text mix (prose, code, chat) on the original weights.

Credits and license

Apache-2.0. See LICENSE.

Downloads last month
-
GGUF
Model size
26B params
Architecture
inkling2
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Harsha001/Inkling-2-Mini-ROCmFPX

Quantized
(40)
this model