Nex-N2.6-oQ3e-mtp

oQ3e mixed-precision MLX conversion of Nex-AGI's Nex-N2.6, intended for Apple Silicon Macs with 192 GB or more of unified memory, including M2 Ultra Mac Studio systems.

Effective weight storage: 3.69 BPW ยท 148.35 GB ยท group size 64.

Nex-N2.6 is a multimodal agentic model based on GLM-5.3-Flash, with approximately 321B parameters. This community conversion retains the full expert set, vision encoder and multi-token-prediction head. No additional training, abliteration or expert pruning was applied.

Conversion

  • Source: creator FP8 checkpoint, nex-agi/Nex-N2.6@987472298e7867721576ac49a253a6f5b526013e.
  • Format: MLX safetensors; oQ3e mixed precision, affine packing, group size 64.
  • Tooling: oMLX 0.7.4.dev1 native oQe streaming converter.
  • Calibration: 256 samples, 512 tokens each, using the bundled code/multilingual corpus; expert-coverage and structural checks passed.
  • Weights: 29 shards, 148.35 GB / 138.16 GiB. Runtime and context state need additional memory.

Detailed conversion and preservation records are in CONVERSION.json.

Mixed precision

3.69 effective bits per weight (BPW). The packed quantized matrices alone average 3.163 bits per weight; the effective figure includes their BF16 scales/biases, every unquantized tensor and safetensors headers.

Component Weight precision Original parameters Stored GB
Routed MoE experts 3-bit, 4-bit 311,653,564,416 137.254404
Attention projections 8-bit 6,171,918,336 6.557663
Shared experts 8-bit 1,082,130,432 1.149764
Dense MLP projections 8-bit 452,984,832 0.481296
Token embeddings 8-bit 634,388,480 0.674038
Output head 8-bit 634,388,480 0.674038
Vision encoder BF16 563,627,008 1.127254
Linear-attention convolutions BF16 3,342,336 0.006685
Indexer pooling/compression BF16 6,297,600 0.012595
Text/MTP normalization BF16 425,216 0.000850
MTP fusion projection BF16 33,554,432 0.067109
MoE routers/correction biases FP32 50,737,248 0.202949
Hyperconnections FP32 35,391,870 0.141567
Recurrent decay controls FP32 280,704 0.001123
Safetensors headers โ€” โ€” 0.000402
Total 3.69 effective BPW 321,323,031,390 148.351738

Stored GB uses decimal units. Quantized rows include their BF16 affine scales/biases; normalization in this table excludes the vision encoder, whose tensors are all counted in the vision row. Routed/shared expert and attention totals include MTP; the MTP fusion projection is listed separately. No tensors are counted twice.

Calculation: 8 ร— 148,351,737,770 stored bytes รท 321,323,031,390 original scalar parameters = 3.6935 BPW. Affine metadata contributes storage but is not counted as additional original model parameters. Tokenizer/config files, KV cache and runtime workspace are excluded.

Per-block precision map
Block Routed expert matrices Other quantized matrices Unquantized tensors Stored GB
Text 0 โ€” (dense MLP) 8-bit BF16, FP32 0.310052
Text 1 โ€” (dense MLP) 8-bit BF16, FP32 0.310052
Text 2 โ€” (dense MLP) 8-bit BF16, FP32 0.310052
Text 3 3-bit 8-bit BF16, FP32 3.338730
Text 4 3-bit 8-bit BF16, FP32 3.351972
Text 5 3-bit 8-bit BF16, FP32 3.351972
Text 6 3-bit 8-bit BF16, FP32 3.351972
Text 7 3-bit 8-bit BF16, FP32 3.338730
Text 8 3-bit 8-bit BF16, FP32 3.351972
Text 9 3-bit 8-bit BF16, FP32 3.351972
Text 10 3-bit 8-bit BF16, FP32 3.351972
Text 11 3-bit 8-bit BF16, FP32 3.338730
Text 12 3-bit 8-bit BF16, FP32 3.351972
Text 13 3-bit 8-bit BF16, FP32 3.351972
Text 14 3-bit 8-bit BF16, FP32 3.351972
Text 15 3-bit 8-bit BF16, FP32 3.338730
Text 16 3-bit 8-bit BF16, FP32 3.351972
Text 17 3-bit 8-bit BF16, FP32 3.351972
Text 18 3-bit 8-bit BF16, FP32 3.351972
Text 19 3-bit 8-bit BF16, FP32 3.338730
Text 20 3-bit 8-bit BF16, FP32 3.351972
Text 21 3-bit 8-bit BF16, FP32 3.351972
Text 22 3-bit 8-bit BF16, FP32 3.351972
Text 23 3-bit 8-bit BF16, FP32 3.338730
Text 24 3-bit 8-bit BF16, FP32 3.351972
Text 25 3-bit 8-bit BF16, FP32 3.351972
Text 26 3-bit 8-bit BF16, FP32 3.351972
Text 27 3-bit 8-bit BF16, FP32 3.338730
Text 28 3-bit 8-bit BF16, FP32 3.351972
Text 29 3-bit 8-bit BF16, FP32 3.351972
Text 30 3-bit 8-bit BF16, FP32 3.351972
Text 31 3-bit 8-bit BF16, FP32 3.338730
Text 32 3-bit 8-bit BF16, FP32 3.351972
Text 33 3-bit 8-bit BF16, FP32 3.351972
Text 34 3-bit 8-bit BF16, FP32 3.351972
Text 35 3-bit 8-bit BF16, FP32 3.338730
Text 36 3-bit 8-bit BF16, FP32 3.351972
Text 37 3-bit 8-bit BF16, FP32 3.351972
Text 38 3-bit 8-bit BF16, FP32 3.351972
Text 39 3-bit 8-bit BF16, FP32 3.338730
Text 40 3-bit 8-bit BF16, FP32 3.351972
Text 41 3-bit 8-bit BF16, FP32 3.351972
Text 42 3-bit 8-bit BF16, FP32 3.351972
Text 43 3-bit 8-bit BF16, FP32 3.338730
Text 44 3-bit 8-bit BF16, FP32 3.351972
MTP 0 4-bit 8-bit BF16, FP32 4.308687

Complete per-module/tensor precision data (CSV) and exact accounting/methodology (JSON) are included. The CSV covers every quantized module and unquantized tensor; quantized module rows include their scales and affine biases.

Running

Use a runtime supporting Nex-N2.6's glm5_next architecture and oQe mixed precision, with the included tokenizer, processor and chat template. The template supports reasoning_effort values low, high and max (default).

Cold loading, text generation, supplied-image recognition and executable coding checks tested on a 256 GB M5 Ultra Mac with oMLX 0.7.4.dev1; MTP decoding active. The generated Dijkstra implementation passed 17 executable checks. Sampled process footprint peaked at 141.81 GiB across the cold load and bounded requests; the largest completed prompt-plus-output was 6,938 tokens.

Intended hardware is a capacity target, not an M2 qualification. M2 Ultra execution, speed and usable context are untested. Runtime, macOS, KV state, generation workspace and the device's Metal/wired-memory budget need headroom beyond the weight files. These are bounded functional checks, not a general quality or full-context benchmark.

Other precision variants

  • Nex-N2.6 oQ4e: 4.64 effective BPW, 186.40 GB.
  • Nex-N2.6 oQ5e: 5.61 effective BPW, 225.36 GB. All listed variants are converted directly from the same pinned creator FP8 checkpoint, not requantized from one another.

Credits and license

Model: Nex-AGI. GLM base: Z.ai. Conversion tooling: oMLX / oQe. Community conversion: Hygaard.

Licensed under Apache-2.0, with the upstream GLM MIT notice retained. See NOTICE for attribution and modification notes.

Downloads last month
123
Safetensors
Model size
321B params
Tensor type
U32
ยท
F32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

3-bit

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

Model tree for Hygaard/Nex-N2.6-oQ3e-mtp

Base model

nex-agi/Nex-N2.6
Quantized
(8)
this model