DEV.co
Open-Source LLM · janhq

Jan-v3.5-4B-gguf

Jan-v3.5-4B is a 4-billion-parameter open-source LLM fine-tuned for math reasoning and conversational personality. It runs locally via vLLM or llama.cpp, uses Apache 2.0 licensing, and targets users who want a smaller model with distinct personality traits rather than neutral assistant behavior.

Source: HuggingFace — huggingface.co/janhq/Jan-v3.5-4B-gguf
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
305.5k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerjanhq
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads305.5k
Likes26
Last updated2026-03-24
Sourcejanhq/Jan-v3.5-4B-gguf

What Jan-v3.5-4B-gguf is

Based on Qwen3-4B architecture with 36 layers, 32 Q-heads and 8 KV-heads (GQA), 3.6B non-embedding parameters. Native context length 262,144 tokens. Fine-tuned on curated math and identity datasets by Menlo Research. Available in GGUF quantization format. No information on original pretraining data, exact parameter counts per variant, or benchmark metrics.

Quickstart

Run Jan-v3.5-4B-gguf locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="janhq/Jan-v3.5-4B-gguf")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local math-focused chatbot

Deploy Jan-v3.5-4B on-device for math reasoning without cloud dependency. Suitable for educational tools, tutoring systems, or technical assistance where math accuracy matters more than scale.

Privacy-first conversational AI

Run entirely on customer infrastructure. Minimal VRAM footprint (~8–12 GB estimated) makes it feasible for workstations, edge servers, or resource-constrained environments needing compliance.

Fine-tuning base for domain-specific models

Use as starting point for downstream math-heavy or identity-specialized applications. Card explicitly encourages this use case.

Running & fine-tuning it

Estimated 8–12 GB VRAM for inference at FP16 or Q8 quantization; lower with Q4/Q5 variants (4–6 GB estimated). Running at FP32 not practical for consumer hardware. CPU-only inference feasible but slow; GPU (NVIDIA/AMD/Metal) strongly recommended. No official specs provided—verification with actual deployment required.

Model card does not discuss LoRA, QLoRA, or parameter-efficient fine-tuning feasibility. As a 4B fine-tuned variant, LoRA adapters are plausible but require separate testing. Full fine-tuning would demand significant VRAM (>16 GB). Contact janhq for official guidance or consult base model Jan-v3-4B-base-instruct documentation.

When to avoid it — and what to weigh

  • Production-scale deployments requiring sub-100ms latency — 4B models incur inference latency unsuitable for high-throughput APIs. Consider larger-capacity serving infrastructure or smaller distilled models.
  • Neutral, corporate-friendly tone required — Jan-v3.5 has deliberate personality (lowercase defaults, casual humor, self-aware style). Not suitable for formal customer service, compliance-heavy interactions, or risk-averse brand voice.
  • Benchmark-driven model selection — No published benchmarks, comparative metrics, or evaluation against standard LLM leaderboards provided. Cannot verify math reasoning claims quantitatively.
  • Extensive code generation or complex reasoning — Training data focuses on math and identity. No mention of code datasets, reasoning chains, or logic-heavy capabilities. Unsuitable for software engineering tasks without validation.

License & commercial use

Apache 2.0 license. Permissive OSI-approved license allowing commercial use, modification, and distribution with attribution and liability disclaimer.

Apache 2.0 is a permissive open-source license that explicitly permits commercial use, derivative works, and private modification. No gating, no model card restrictions beyond attribution. Commercial deployment is legally supported. However, downstream liability and warranty disclaimers apply per Apache 2.0 terms. Verify compliance with your legal counsel if model outputs or behavior carry material risk.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityLow
DEV.co fitGood
Assessment confidenceMedium
Security considerations

No security audit, adversarial robustness data, or harmful output testing documented. Personality training may increase likelihood of sarcasm, humor, or off-brand outputs in sensitive contexts. Recommend sandboxing and output validation for production. Model runs locally (reduces supply-chain risk vs. API-only models). No mention of data poisoning detection or training data filtering.

Alternatives to consider

Llama 2 7B or Mistral 7B

Similar parameter count, more proven benchmarks and broader community adoption. Neutral tone. Better for general-purpose use without personality constraints.

Qwen2.5-4B (if available)

Same family as Jan's base architecture; likely better benchmarks and official Alibaba support. Better for math without personality overhead.

TinyLlama 1.1B or Phi-2

Smaller footprint for edge/mobile. Trade math reasoning for lower latency and VRAM. Only if you prioritize hardware constraints over capability.

Software development agency

Ship Jan-v3.5-4B-gguf with senior software developers

Test Jan-v3.5-4B on your math and conversational workloads. Start with vLLM or llama.cpp, or integrate directly into Jan Desktop. For production sizing, infrastructure planning, or fine-tuning guidance, consult our AI engineering team.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Jan-v3.5-4B-gguf FAQ

Can I use Jan-v3.5-4B in a commercial product?
Yes. Apache 2.0 permits commercial use, modification, and distribution. You must retain the license notice and liability disclaimer. Review your legal counsel for downstream risk management, especially if model outputs carry material consequences.
What GPU do I need?
An NVIDIA GPU with 12 GB+ VRAM (RTX 3060 or better) is recommended for FP16 inference. Smaller quantizations (Q4/Q5) may run on 8 GB. CPU-only inference is possible but slow. Apple Metal and AMD GPUs are supported via llama.cpp.
How does Jan-v3.5 compare to the base Jan-v3-4B model in math performance?
Model card claims 'improved mathematical problem-solving' but provides no quantitative benchmarks, ablations, or before/after metrics. Verify performance on your specific math tasks before production deployment.
Will the personality affect my use case?
Yes. Jan-v3.5 defaults to lowercase, casual tone, self-aware humor, and avoids corporate language. Test outputs on representative prompts. If your application requires formal, neutral tone, consider base models like Mistral or Llama 2 instead.

Work with a software development agency

Need help beyond evaluating Jan-v3.5-4B-gguf? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy Jan-v3.5-4B?

Test Jan-v3.5-4B on your math and conversational workloads. Start with vLLM or llama.cpp, or integrate directly into Jan Desktop. For production sizing, infrastructure planning, or fine-tuning guidance, consult our AI engineering team.