Nex-N2.5 mini · MLX 4bit
A community standard MLX conversion by TensorFold for Apple Silicon.
Model
Converted from the BF16 Nex-N2.5-mini checkpoint using MLX-VLM affine quantisation with 4-bit weights and group size 64. Not every tensor is quantised: the converter reported 4.649 average bits per weight, and multimodal modules excluded by its default predicate retain higher precision. This text-and-vision model uses the qwen3_5_moe architecture. The inspected BF16 checkpoint contained no matching MTP tensors, despite its configuration declaring one MTP layer. This release does not provide tested MTP decoding; keep MTP disabled.
Download and use
hf download TensorFold/Nex-N2.5-mini-MLX-4bit --local-dir ./Nex-N2.5-mini-MLX-4bit
Add the folder to oMLX model directories, refresh the model list and select it. Basic inference was tested with oMLX 0.6.4.
Use the upstream-recommended sampling:
{
"temperature": 0.7,
"top_p": 0.95,
"top_k": 40
}
Set these explicitly in your client or model settings. Our sampled oQ2 retests also used reasoning_effort="none" and max_tokens=4096; these are test conditions, not an upstream recommendation to disable reasoning. Avoid greedy decoding for oQ2, which reproduced a repetition loop in our tests.
Validation and limitations
Tested on 9 September 2026 through oMLX 0.6.4 on an Apple Silicon Studio with 256 GiB unified memory. Exact arithmetic, a forced weather-tool call with a Paris argument, and identification of a synthetic red image passed for this quant. Tool calls were checked for formatting, not executed. These are basic checks, not a full coding, vision or agent evaluation.
Four coding requests were tested: one finished naturally and three hit the 4096-token cap. No repeated-12-word sequence occurred more than twice. Complete code blocks were syntax-checked, not executed; incomplete responses and code correctness remain limitations. The structured JSON test passed. The story finished at 925 words, exceeding the requested 700–900.
The four coding requests delivered approximately 103–104 output tokens per API elapsed second. This is end-to-end throughput, not a controlled decode-only benchmark; cache state was not controlled. Peak request memory and context-fit limits have not been measured, so no Mac memory-tier recommendation is claimed. Long-context, multi-turn and broader vision quality remain unverified.
Short test
With the sampling settings above, try:
What is 17 multiplied by 19? Answer with only the number.
The recorded arithmetic check returned 323 using the recommended sampling above and reasoning_effort="none".
A short correct response does not establish long-generation reliability.
Licence and attribution
The upstream repository declares Apache-2.0. Model training, architecture and the original Nex logo belong to Nex-AGI and the respective upstream contributors. This is an independent community conversion, not an official Nex-AGI release. Upstream benchmark scores are not evaluations of this quant.