All models
Text / ReasoningOpen weights4B total

Vikasit 3.5 4B

Next-gen 4B with improved reasoning and multimodal awareness.

Overview

Vikasit 3.5 4B is the next-generation 4B — a hybrid-attention MoE with markedly improved reasoning and multimodal awareness, 262K native context expandable to ~1M. Strong quality from a small footprint.

Specifications

Total parameters
4B total
Architecture
Hybrid MoE (Gated DeltaNet + sparse MoE)
Layers
32
Context window
262K native, ~1M via YaRN
Modalities
Text in → text out (multimodal-capable base)
License
Apache 2.0

Capabilities

  • Improved reasoning over previous 4B
  • 262K native context, ~1M via YaRN
  • Strong instruction following (IFEval 89.8)
  • Thinking and non-thinking modes
Multilingual. Strong English + major Indian languages.

Benchmarks

BenchmarkScore
MMLU-Pro79.1
GPQA-Diamond76.2
LiveCodeBench v655.8
IFEval89.8
HMMT Feb 202574.0
MATH-500N/A

Built on Qwen3.5-4B; numbers from the Qwen3.5-4B HuggingFace model card. HMMT reported instead of AIME 2025.

Hardware & deployment

PrecisionMemory
bf16~8 GB
INT4~2.5 GB

Quick start

Vikasit 3.5 4B is an open-weight model. Self-host it with any OpenAI-compatible inference server and call it with the OpenAI SDK as shown below.

OpenAI-compatible Python (self-hosted, e.g. vLLM)
# pip install openai
import os
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="sk-local",  # self-hosted servers accept any token
)

resp = client.chat.completions.create(
    model="vikasit-3.5-4b",
    messages=[
        {"role": "user", "content": "Explain Vikasit 3.5 4B in one sentence."}
    ],
)

print(resp.choices[0].message.content)

Limitations

  • Hybrid-attention kernels need recent runtimes
  • Some classic benchmarks not published by base

Vikasit 3.5 4B FAQ

How much does Vikasit 3.5 4B cost?

Vikasit 3.5 4B is an open-weight model built on Qwen3.5-4B (Apache 2.0). Self-hosting the weights is free under the Apache 2.0 licence — you pay only for the hardware or cloud GPUs you run it on. Typical deployment fits the memory profiles listed in the hardware section above.

Is Vikasit 3.5 4B open weight?

Yes. Vikasit 3.5 4B is built on Qwen3.5-4B (Apache 2.0) and distributed under the Apache 2.0 licence, so the weights are openly available for self-hosting, fine-tuning, and commercial use, subject to the upstream licence terms.

How do I run Vikasit 3.5 4B?

Because Vikasit 3.5 4B is open weight, you self-host it with any OpenAI-compatible inference server (such as vLLM or SGLang) loaded with the Qwen3.5-4B (Apache 2.0) weights, then call it with the OpenAI SDK by setting the base URL to your own endpoint.

What context window does Vikasit 3.5 4B support?

Vikasit 3.5 4B supports a 262K native, ~1M via YaRN context window. It is a 4B total Hybrid MoE (Gated DeltaNet + sparse MoE) model — full specifications are listed in the table above.

License & attribution

Apache 2.0

Built on Qwen3.5-4B (Apache 2.0). Upstream copyright, license, and attribution notices are retained.

Qwen3.5-4B model card