Nex-N2.5-mini-NVFP4

A plain, data-free NVFP4 quantization of nex-agi/Nex-N2.5-mini, produced by ProCreations. The checkpoint on main is the plain version evaluated in the results below.

Packed weight files occupy 23.93 GB, compared with 70.21 GB for the BF16 source: 65.9% smaller. These are decimal weight-file sizes; total serving VRAM also includes runtime buffers and context storage.

Quantization

Expert weights use standard maximum-based NVFP4 scaling. Expert activation global input scales are fixed at 1.0, while per-block activation scales remain dynamic. This adapts NVIDIA's experts-only input-scale-1 recipe, with its FP8 KV-cache option omitted.

The conversion used zero calibration examples and zero model calibration forwards. No fine-tuning or Hessian/MSE scale search was performed for this checkpoint. ModelOpt export separately ran a two-token structural probe with quantization and calibration-statistics collection disabled. All 5,789,854,432 bytes of nonexpert source tensors were verified to retain their original dtype and exact values. Original tokenizer, chat-template and processor files are preserved.

Evaluation

The published weights were evaluated with native W4A4 NVFP4 kernels on an RTX PRO 6000, using the same prompts, runtime and scoring as the pinned BF16 source. These are 472 diagnostic cases plus 12 synthetic long tool workflows, not official full-benchmark results.

The equal-weight mean of the eight broad diagnostic suites is 90.87% for BF16 and 90.30% for plain NVFP4. BFCL totals are 48/96 and 45/96, respectively. No invalid BFCL tool calls were recorded in either model's tested cases. The synthetic workflows reached 225,650 actual prompt tokens with plain NVFP4, with no invalid calls or truncated steps.

Broad reasoning tests used temperature 0 and an 8,192-token output limit; BFCL used 4,096 tokens per step. Truncated answers remain included. The synthetic long workflows use archives around 32K, 98K and 224K tokens and dependent tool calls. They do not replace real repository coding-agent or browser/computer-use benchmarks. Teacher-distribution KL and top-1 fidelity were not measured for this plain checkpoint. Small diagnostic differences do not establish quality retention across every workload.

Full per-case results, generation limits and package/data revisions: native_report.json, long_agent_report.json. Evaluation sources include GSM8K, MMLU-Pro, DocumentVQA, HumanEval+, LongBench v2, and BFCL.

Earlier MSE and local-Hessian variants were measured on the same native suite. Local Hessian averaged 90.36% broadly and passed 46/96 BFCL cases, versus 90.30% and 45/96 for plain; plain passed more of the long-context BFCL subset (21/32 versus 17/32). All variants passed 12/12 synthetic workflows. See the comparison data; older experiments are retained for reference and do not describe the weights on main.

Inference

Validated environment: SGLang 0.5.18, PyTorch 2.13.0+cu130, Transformers 5.12.1, using flashinfer_cutlass for both MoE execution and FP4 GEMM on Blackwell hardware. The exact tested container is:

nexagi/sglang@sha256:7b788151280100898b345452bcb88131d83d054bc5a9ae7d15f9d1f85eba60ac

Inside that environment, download the repository and launch with the tested settings:

hf download ProCreations/Nex-N2.5-mini-NVFP4 --local-dir ./Nex-N2.5-mini-NVFP4
python -m sglang.launch_server \
  --model-path ./Nex-N2.5-mini-NVFP4 \
  --served-model-name nvfp4 \
  --host 127.0.0.1 \
  --port 31001 \
  --tp-size 1 \
  --context-length 262144 \
  --max-running-requests 4 \
  --chunked-prefill-size 4096 \
  --mem-fraction-static 0.88 \
  --cuda-graph-max-bs 4 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --chat-template ./Nex-N2.5-mini-NVFP4/chat_template.jinja \
  --mamba-scheduler-strategy extra_buffer \
  --moe-runner-backend flashinfer_cutlass \
  --fp4-gemm-backend flashinfer_cutlass

The example listens on localhost at port 31001 and uses the tested context limit of 262,144 tokens. Memory requirements depend on context and concurrency. The preserved source chat template supports the original reasoning and tool-use controls.

Provenance

Source revision: 87420286149d9cce9bd46cd335ef9bda33c37c1b. NVIDIA Model Optimizer revision: 5cae3940402f1ced98069a666b0bec72ec8b33b5. The pinned source archive identifies its package version as 0.0.0; the commit identifies the actual build.

Build and file-hash provenance, export and evaluation status, plain conversion script, and environment setup are included. Every inference file was content-hash verified before evaluation, and the files on main match those tested weights and configuration.

Released under Apache 2.0; see NOTICE. This is an independent ProCreations quantization. The source model's capabilities and limitations still apply.