DEV.co
Open-Source LLM · Menlo

Jan-nano-128k

Jan-Nano-128k is a 4B-parameter open-source language model from Menlo Research with a native 128k token context window, designed for research and document analysis. It builds on the Jan-Nano base model and is distributed under Apache 2.0. The model is non-gated and can be self-hosted via vLLM, llama.cpp, or similar inference servers.

Source: HuggingFace — huggingface.co/Menlo/Jan-nano-128k
4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
59.6k
Downloads (30d)

Key facts

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

FieldValue
DeveloperMenlo
Parameters4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads59.6k
Likes223
Last updated2025-07-01
SourceMenlo/Jan-nano-128k

What Jan-nano-128k is

A 4.02B-parameter transformer-based text-generation model fine-tuned from Qwen3 architecture with extended RoPE (rope_type: yarn, factor 3.2, original_max_position_embeddings 40960). Natively supports 128k context length. Last updated 2025-07-01. Inference tested with vLLM and llama.cpp; compatible with text-generation-inference endpoints. Chat template included in tokenizer (Qwen3 non-thinking format).

Quickstart

Run Jan-nano-128k locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Menlo/Jan-nano-128k")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

Long-form Research Document Analysis

Process entire research papers, technical documentation, and lengthy reports in a single context window without truncation or rolling buffers. Ideal for systematic literature reviews and comparative analysis.

Multi-turn Conversation with Full Context Retention

Maintain comprehensive conversation history and reference earlier exchanges without performance degradation across the full 128k window. Suitable for deep research dialogues and collaborative problem-solving.

Self-Hosted Private Inference

Deploy locally via vLLM or llama.cpp on modest hardware (estimated 8–16 GB VRAM) for document processing without external API calls or data egress concerns.

Running & fine-tuning it

ESTIMATE: 8–16 GB GPU VRAM for inference at full 128k context (fp16/bfloat16). Exact requirements depend on batch size and serving framework (vLLM with auto-scaling may require upper bound). CPU-only inference via llama.cpp possible but slower. Verify empirically for your infrastructure before production deployment.

Unknown whether LoRA/QLoRA adapters have been published or tested on this model. Model card does not mention fine-tuning recipes or community fine-tuning efforts. Contact developer (GitHub issues or HF discussions) for guidance on custom domain adaptation. Standard transformer fine-tuning tools (transformers library, Axolotl) likely compatible but untested.

When to avoid it — and what to weigh

  • Latency-Critical Real-Time Applications — Processing 128k tokens incurs non-trivial inference latency. Not suitable for sub-second response requirements or high-throughput streaming scenarios.
  • Superior Reasoning or Creative Output — Model is explicitly non-thinking (no internal reasoning chain). May underperform on complex logic, math, or creative tasks compared to larger reasoning-enabled models.
  • Hardware-Constrained Environments — While compact relative to larger LLMs, running full 128k context locally requires 8–16 GB+ VRAM. Not suitable for edge devices, mobile, or heavily resource-constrained setups.
  • Tasks Requiring Domain-Specific Expertise Without Fine-tuning — Base model was not fine-tuned for specialized domains (medical, legal, financial). Domain tasks require additional fine-tuning or RAG augmentation.

License & commercial use

Licensed under Apache 2.0, a permissive OSI-approved license. Permits commercial use, distribution, modification, and private use. Requires retention of copyright notice and license text.

Apache 2.0 explicitly permits commercial use, commercial distribution, and private deployment. No usage restrictions for commercial products or services. However, verify your derivative work (if any) complies with the license. No commercial support, SLA, or indemnification from Menlo Research is stated; support is community-driven via HF discussions and GitHub issues.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit or vulnerability disclosure process documented. As an open-source model, security relies on community review and standard transformer inference safeguards. Self-hosted deployment avoids third-party API/data exposure. Review inference server (vLLM, llama.cpp) security practices separately. No information on prompt injection or jailbreak testing.

Alternatives to consider

Llama 2 / Llama 3 with RoPE extension

Larger parameter counts (7B–70B) with more established community and enterprise support. May offer better performance on complex tasks but higher hardware cost and no native 128k training.

Qwen (full size) models with native long context

Base architecture is Qwen3; larger Qwen variants may offer stronger performance and broader community adoption, though increased deployment complexity and cost.

Mistral 7B with context extension

Larger parameter count (7B vs. 4B) with well-established open-source community. Supports extended context via rope scaling but not natively trained on it.

Software development agency

Ship Jan-nano-128k with senior software developers

Jan-Nano-128k enables research-grade document processing on your infrastructure. Start with our deployment guides for vLLM or llama.cpp, or explore how Devco can help architect a private LLM pipeline for your 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-nano-128k FAQ

Can I use Jan-Nano-128k in a commercial product?
Yes. Apache 2.0 license permits commercial use, distribution, and deployment. No licensing fee or approval required. Ensure you retain the copyright notice and license text in your product or documentation. No commercial support SLA is provided by Menlo Research; community support is available via GitHub and HuggingFace discussions.
What GPU do I need to run this locally?
Estimated 8–16 GB VRAM for fp16/bfloat16 inference at full 128k context, depending on batch size and inference framework. Smaller GPUs (4–6 GB) may work with quantization (int8, int4 via GGUF or AutoGPTQ) or reduced max sequence length. CPU-only inference is possible via llama.cpp but will be slow. Test on your target hardware first.
How does Jan-Nano-128k compare to the original Jan-Nano?
Jan-Nano-128k extends the context window from the base Jan-Nano (40.96k) to 128k tokens using native RoPE scaling (yarn, factor 3.2). Model card claims improved performance across the full context range compared to traditional extension methods. See SimpleQA benchmark chart in model card for empirical comparison.
Is this model suitable for reasoning and math?
Model is explicitly non-thinking, meaning it lacks internal reasoning chains. Performance on complex logic, math, and abstract reasoning may be limited compared to reasoning-enabled models. Best suited for document retrieval, synthesis, and conversational tasks.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Jan-nano-128k is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Long-Context Intelligence Privately?

Jan-Nano-128k enables research-grade document processing on your infrastructure. Start with our deployment guides for vLLM or llama.cpp, or explore how Devco can help architect a private LLM pipeline for your team.