Nex-N2.5-mini (nex-agi) : MoE 35B / ~3B actifs post-trained on Qwen3.5-35B-A3B. Vision, tool-calling, reasoning, 256k context. Quants Q5_K_M, Q4_K_M (latest), Q3_K_M, Q2_K.

Details

Updated 6 days ago

6 days ago

ab5cd99f1095 · 15GB

model

archqwen35moe

parameters34.7B

quantizationQ2_K

14GB

projector

archclip

parameters447M

quantizationF16

899MB

license

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US

12kB

params

{ "repeat_penalty": 1, "temperature": 0.7, "top_k": 40, "top_p": 0.95 }

63B

Nex-N2.5-mini

Vision · Tools · Thinking · 262K context · Apache-2.0

Nex-N2.5-mini is the small member of nex-agi’s Nex-N2.5 agentic model family, post-trained from Qwen3.5-35B-A3B. Sparse MoE: 35B total / ~3B active parameters, so it decodes far faster than its size suggests.

It is built for long-horizon agentic work: computer use, web browsing, tool calling, and visually grounded tasks. Vision is not just an input modality here — it is the feedback channel the model uses to check its own work.

GGUF quantizations by bartowski · original weights by nex-agi

Tag	Quant	Size	Notes
`latest`	Q4_K_M	22.3 GB	best size/quality compromise, fits a 24 GB GPU
`q5_k_m`	Q5_K_M	27.0 GB	high quality, 32 GB GPU
`q3_k_m`	Q3_K_M	17.3 GB	good for 16 GB GPUs
`q2_k`	Q2_K	13.8 GB	smallest, quality visibly degraded

Usage

# chat
ollama run tobestyledintro/nex-n2.5-mini

# long context (native 262144, raise it explicitly)
ollama run tobestyledintro/nex-n2.5-mini --ctx 131072

# vision: drag an image, or
ollama run tobestyledintro/nex-n2.5-mini "Read the invoice in ./facture.png and give me the total"

API:

curl http://localhost:11434/api/chat -d '{
  "model": "tobestyledintro/nex-n2.5-mini",
  "messages": [{"role": "user", "content": "Hello!"}]
}'

Python (vision):

import ollama
r = ollama.chat(
    model='tobestyledintro/nex-n2.5-mini',
    messages=[{'role': 'user', 'content': 'What is in this image?',
               'images': ['screenshot.png']}],
)
print(r.message.content)

Capabilities

Vision — image understanding (clip projector, 447M params).
Tool calling — native, handled by Ollama’s qwen3.5 parser: proper structured tool_calls, no XML to parse yourself.
Thinking — reasoning trace in <think> tags, handled by Ollama’s qwen3.5 renderer. Nex-N2.5 exposes three modes through reasoning_effort: "none" (answer directly), "medium" (default, adaptive), "high" (always think).
Context — 262144 tokens natively; the KV cache is small thanks to the hybrid linear/full attention stack (only 10 of 40 layers are full attention), so long contexts stay cheap.

Recommended sampling

nex-agi’s own recommendation, used as defaults in this model: temperature=0.7, top_p=0.95, top_k=40.

Performance

Measured on 1× RTX 5090 (32 GB), Q4_K_M fully offloaded (OLLAMA_FLASH_ATTENTION=1, OLLAMA_KV_CACHE_TYPE=q8_0):


model load	~8.6 s (22 GB to VRAM)
generation	116–163 tok/s
prompt processing	5–10k tok/s up to ~25k tokens
generation @ 120k context	~90 tok/s
generation @ 250k context	~74 tok/s

No speculative decoding. The upstream config declares an MTP head (mtp_num_hidden_layers: 1) but no MTP tensors are shipped in the released weights, so DRAFT / draft_num_predict / --spec-type draft-mtp do not apply. The speeds above are the ceiling.

Architecture


architecture	`qwen35moe` (Qwen3.5 MoE)
layers	40 (30 linear attention + 10 full attention)
experts	256 per layer, 8 active
hidden size	2048
heads	16 Q / 2 KV, head_dim 256
vision	clip `qwen3vl_merger`, 27 blocks, 1152 → 2048
vocab	248320
quantization	Q4_K_M via imatrix (bartowski)

Credit & license

Original model: nex-agi/Nex-N2.5-mini — Apache-2.0
GGUF quants: bartowski
This Ollama packaging: community repack, Apache-2.0. All credit to nex-agi.