JetLLMLite-3.6

This repository hosts an organization-managed copy of JetLLMLite-3.6 for advanced coding, reasoning, long-context, and agentic AI workloads.

It is intended for teams that want to manage deployment, access, and internal distribution from their own namespace while preserving compatibility with the upstream model ecosystem.

Model Summary

JetLLMLite-3.6 is an open-weight post-trained model released in Hugging Face Transformers format. According to the official model card, these artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, and KTransformers. The model is described as a Causal Language Model with Vision Encoder, with 35B total parameters and 3B activated parameters, plus 262,144 native context length extensible up to 1,010,000 tokens. :contentReference[oaicite:0]{index=0}

Key Features

  • 35B total parameters
  • 3B activated parameters
  • MoE-based architecture
  • Vision-language capability
  • Strong coding and agentic performance
  • 262,144 native context length
  • Extensible context up to 1,010,000 tokens
  • Compatible with Transformers, vLLM, SGLang, and KTransformers :contentReference[oaicite:1]{index=1}

Intended Use

This model is suitable for:

  • advanced chat assistants
  • coding assistants
  • repository-level reasoning
  • agentic workflows
  • multimodal question answering
  • long-context document understanding
  • RAG and tool-using systems
  • enterprise AI applications
  • research and benchmarking :contentReference[oaicite:2]{index=2}

Model Details

Architecture

According to the official Qwen model card:

  • Model type: Causal Language Model with Vision Encoder
  • Training stage: Pre-training & Post-training
  • Total parameters: 35B
  • Activated parameters: 3B
  • Hidden dimension: 2048
  • Token embedding: 248320 (padded)
  • Number of layers: 40
  • Hidden layout: 10 ร— (3 ร— (Gated DeltaNet โ†’ MoE) โ†’ 1 ร— (Gated Attention โ†’ MoE))
  • Number of experts: 256
  • Activated experts: 8 Routed + 1 Shared
  • Expert intermediate dimension: 512
  • MTP: trained with multi-steps
  • Context length: 262,144 natively and extensible up to 1,010,000 tokens :contentReference[oaicite:3]{index=3}

Performance Positioning

The official model card highlights JetLLMLite-3.6 as an open-weight model built with a focus on stability and real-world utility, with notable emphasis on:

  • Agentic coding
  • Frontend workflows
  • Repository-level reasoning
  • Thinking preservation across historical context :contentReference[oaicite:4]{index=4}

Hardware Requirements

This model does not have a single universal minimum hardware requirement for every deployment scenario.

Actual requirements depend on:

  • inference backend
  • precision / quantization
  • batch size
  • context length
  • whether vision inputs are enabled
  • concurrency
  • latency targets
  • KV cache configuration :contentReference[oaicite:5]{index=5}

Minimum System Requirements

Because the model has 35B total parameters and 3B activated parameters, real VRAM usage can vary substantially depending on the runtime and workload profile. The upstream model card does not publish a single hard minimum VRAM number. However, for practical planning, the following guidance is reasonable:

  • Estimated practical minimum for heavily quantized local inference: around 24 GB VRAM
  • More realistic for smoother local / development usage: 48โ€“80 GB VRAM
  • Recommended for production serving of the original model: multi-GPU or high-memory datacenter GPU environments
  • Recommended for long-context or multimodal serving: high-memory datacenter-class infrastructure :contentReference[oaicite:6]{index=6}

Note: these values are practical estimates for deployment planning, not universal hard limits. Real memory usage can increase significantly with longer contexts, multimodal inputs, larger batch sizes, and serving-framework overhead.

Reference Hardware

The upstream model card provides official serving examples for this model with common inference stacks such as vLLM, SGLang, and Transformers, and includes a Docker example for SGLang. For practical deployment planning:

  • Quantized local experimentation: high-memory single-GPU environments may be sufficient
  • Standard production-oriented serving: modern datacenter GPUs are recommended
  • Long-context and higher-concurrency serving: multi-GPU deployment is the safer reference setup
  • Multimodal production workloads: high-memory server infrastructure is strongly recommended :contentReference[oaicite:7]{index=7}

Practical Recommendation

For most teams:

  1. start with quantized evaluation if you are testing locally
  2. benchmark using your real context lengths
  3. use dedicated serving stacks such as vLLM or SGLang
  4. reserve high-memory infrastructure for production-scale or long-context workloads :contentReference[oaicite:8]{index=8}

Software Requirements

Recommended environment:

  • Python 3.10+
  • Linux
  • CUDA-enabled GPU environment
  • One of the following runtimes:
    • Transformers
    • vLLM
    • SGLang
    • KTransformers

The official model card also notes that the latest transformers is required for JetLLMLite-3.6 and recommends ensuring torchvision and pillow are installed for multimodal use. :contentReference[oaicite:9]{index=9}

Common dependencies may include:

  • torch
  • transformers
  • torchvision
  • pillow

Quickstart

Install Transformers:

pip install "transformers[serving]"

Basic loading example:

from transformers import pipeline

pipe = pipeline(
    "image-text-to-text",
    model="Jetlink/JetLLMLite-3.6",
    trust_remote_code=True
)

Serving Examples

vLLM

vllm serve Jetlink/JetLLMLite-3.6

SGLang

docker run --gpus all \
  --shm-size 32g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=<secret>" \
  --ipc=host \
  lmsysorg/sglang:latest \
  python3 -m sglang.launch_server \
      --model-path "Jetlink/JetLLMLite-3.6" \
      --host 0.0.0.0 \
      --port 30000

Docker Model Runner

docker model run hf.co/Jetlink/JetLLMLite-3.6

Long Context Notes

JetLLMLite-3.6 natively supports 262,144 tokens and can be extended up to 1,010,000 tokens. The official model card includes YaRN-style rope scaling configuration guidance for long-context usage. :contentReference[oaicite:10]{index=10}

Strengths

  • strong coding and agentic capabilities
  • improved frontend and repository-level reasoning
  • multimodal support
  • very long native context
  • extensible ultra-long context support
  • modern MoE architecture
  • compatibility with popular open inference frameworks :contentReference[oaicite:11]{index=11}

Limitations

  • infrastructure requirements can be substantial depending on deployment style
  • long-context inference can greatly increase memory pressure
  • multimodal workloads add additional overhead
  • local deployment practicality depends heavily on quantization and runtime choices
  • real-world latency and throughput vary significantly by framework and hardware configuration

Out-of-Scope / Cautionary Use

Outputs should be reviewed before use in:

  • medical decision-making
  • legal advice
  • safety-critical automation
  • high-stakes financial decisions
  • fully autonomous actions without validation
  • sensitive production workflows without guardrails

Human review, tool validation, and policy controls are strongly recommended.

License

This repository follows the same license as the upstream release.

  • License: Apache-2.0

If you redistribute, fine-tune, quantize, or otherwise modify this model, make sure your usage remains compliant with the upstream license and attribution requirements. :contentReference[oaicite:12]{index=12}

Attribution

Original upstream model:

  • Qwen/Qwen3.6-35B-A3B

This repository is an organization-managed copy and is not the original upstream source.

Citation

Please cite the original JetLLMLite-3.6 release when using this model in research, evaluation, or production documentation.

Disclaimer

This repository may include packaging, naming, or deployment-oriented changes for organizational use.

For official updates, benchmark details, long-context settings, and upstream release notes, refer to the original Qwen model card. :contentReference[oaicite:13]{index=13}


JetLLMLite-3.6 (Tรผrkรงe)

Bu depo, geliลŸmiลŸ kodlama, akฤฑl yรผrรผtme, uzun baฤŸlam ve agentic AI iลŸ yรผkleri iรงin JetLLMLite-3.6 modelinin kurum tarafฤฑndan yรถnetilen bir kopyasฤฑnฤฑ barฤฑndฤฑrฤฑr.

Bu depo; modeli kendi namespaceโ€™i altฤฑnda yรถnetmek, eriลŸimi kontrol etmek ve daฤŸฤฑtฤฑmฤฑ kolaylaลŸtฤฑrmak isteyen ekipler iรงin hazฤฑrlanmฤฑลŸtฤฑr. Amaรง, upstream model ekosistemiyle uyumluluฤŸu koruyarak kurumsal kullanฤฑm saฤŸlamaktฤฑr.

Model ร–zeti

JetLLMLite-3.6 resmi model kartฤฑna gรถre bu artefaktlar Hugging Face Transformers, vLLM, SGLang ve KTransformers ile uyumludur. Model, Vision Encoder iรงeren bir Causal Language Model olarak tanฤฑmlanฤฑr; 35B toplam parametre, 3B aktif parametre, 262.144 token yerel baฤŸlam ve 1.010.000 tokenโ€™a kadar geniลŸletilebilir baฤŸlam sunar. :contentReference[oaicite:14]{index=14}

Temel ร–zellikler

  • 35B toplam parametre
  • 3B aktif parametre
  • MoE tabanlฤฑ mimari
  • Vision-language kabiliyeti
  • Gรผรงlรผ kodlama ve agentic performans
  • 262.144 token yerel baฤŸlam
  • 1.010.000 tokenโ€™a kadar geniลŸletilebilir baฤŸlam
  • Transformers, vLLM, SGLang ve KTransformers ile uyumluluk :contentReference[oaicite:15]{index=15}

Kullanฤฑm Amacฤฑ

Bu model aลŸaฤŸฤฑdaki senaryolar iรงin uygundur:

  • geliลŸmiลŸ sohbet asistanlarฤฑ
  • kodlama asistanlarฤฑ
  • repository seviyesinde akฤฑl yรผrรผtme
  • agentic workflow yapฤฑlarฤฑ
  • multimodal soru-cevap
  • uzun baฤŸlamlฤฑ dokรผman anlama
  • RAG ve tool-using sistemler
  • kurumsal AI uygulamalarฤฑ
  • araลŸtฤฑrma ve benchmark รงalฤฑลŸmalarฤฑ :contentReference[oaicite:16]{index=16}

Model Detaylarฤฑ

Mimari

Resmi Qwen model kartฤฑna gรถre:

  • Model tipi: Vision Encoder iรงeren Causal Language Model
  • EฤŸitim aลŸamasฤฑ: Pre-training & Post-training
  • Toplam parametre: 35B
  • Aktif parametre: 3B
  • Hidden dimension: 2048
  • Token embedding: 248320 (padded)
  • Katman sayฤฑsฤฑ: 40
  • Hidden layout: 10 ร— (3 ร— (Gated DeltaNet โ†’ MoE) โ†’ 1 ร— (Gated Attention โ†’ MoE))
  • Expert sayฤฑsฤฑ: 256
  • Aktif expert: 8 Routed + 1 Shared
  • Expert intermediate dimension: 512
  • MTP: multi-steps ile eฤŸitilmiลŸ
  • BaฤŸlam uzunluฤŸu: yerel olarak 262.144, geniลŸletilmiลŸ olarak 1.010.000 tokenโ€™a kadar :contentReference[oaicite:17]{index=17}

Performans Konumlandฤฑrmasฤฑ

Resmi model kartฤฑ JetLLMLite-3.6โ€™yฤฑ รถzellikle ลŸu alanlarda รถne รงฤฑkarฤฑr:

  • Agentic coding
  • Frontend workflowโ€™leri
  • Repository-level reasoning
  • GeรงmiลŸ mesajlardan dรผลŸรผnme baฤŸlamฤฑnฤฑ koruma :contentReference[oaicite:18]{index=18}

Donanฤฑm Gereksinimleri

Bu model iรงin her senaryoya uyan tek bir evrensel minimum donanฤฑm gereksinimi yoktur.

Gerรงek ihtiyaรงlar ลŸunlara baฤŸlฤฑdฤฑr:

  • inference backend
  • precision / quantization
  • batch size
  • baฤŸlam uzunluฤŸu
  • vision girdilerinin aรงฤฑk olup olmamasฤฑ
  • concurrency
  • latency hedefleri
  • KV cache yapฤฑlandฤฑrmasฤฑ :contentReference[oaicite:19]{index=19}

Minimum Sistem Gereksinimleri

Model 35B toplam parametreye ve 3B aktif parametreye sahip olduฤŸu iรงin gerรงek VRAM kullanฤฑmฤฑ runtime ve iลŸ yรผkรผne gรถre ciddi biรงimde deฤŸiลŸebilir. Upstream model kartฤฑ tek bir kesin minimum VRAM sayฤฑsฤฑ vermez. Ancak pratik planlama iรงin ลŸu rehber uygundur:

  • AฤŸฤฑr quantized local inference iรงin tahmini pratik minimum: yaklaลŸฤฑk 24 GB VRAM
  • Daha rahat local / geliลŸtirme kullanฤฑmฤฑ iรงin daha gerรงekรงi seviye: 48โ€“80 GB VRAM
  • Orijinal modelin production servingโ€™i iรงin รถnerilen: รงoklu GPU veya yรผksek bellekli datacenter GPU ortamlarฤฑ
  • Uzun baฤŸlam veya multimodal serving iรงin รถnerilen: yรผksek bellekli datacenter sฤฑnฤฑfฤฑ altyapฤฑ :contentReference[oaicite:20]{index=20}

Not: bunlar deployment planlamasฤฑ iรงin pratik tahminlerdir; evrensel kesin sฤฑnฤฑrlar deฤŸildir. Daha uzun baฤŸlamlar, multimodal girdiler, bรผyรผk batch size ve serving framework kaynaklฤฑ ek yรผkler gerรงek bellek kullanฤฑmฤฑnฤฑ ciddi ลŸekilde artฤฑrabilir.

Referans Donanฤฑm

Upstream model kartฤฑ bu model iรงin vLLM, SGLang ve Transformers gibi inference stackโ€™leriyle resmi serving รถrnekleri sunar ve ayrฤฑca SGLang iรงin bir Docker รถrneฤŸi iรงerir. Pratik daฤŸฤฑtฤฑm planlamasฤฑ iรงin:

  • Quantized local denemeler: yรผksek bellekli tek GPU ortamlarฤฑ yeterli olabilir
  • Standart production serving: modern datacenter GPUโ€™lar รถnerilir
  • Uzun baฤŸlam ve daha yรผksek concurrency: รงoklu GPU daฤŸฤฑtฤฑmฤฑ daha gรผvenli referanstฤฑr
  • Multimodal production iลŸ yรผkleri: yรผksek bellekli sunucu altyapฤฑsฤฑ gรผรงlรผ ลŸekilde รถnerilir :contentReference[oaicite:21]{index=21}

Pratik ร–neri

ร‡oฤŸu ekip iรงin en mantฤฑklฤฑ yaklaลŸฤฑm:

  1. local testte quantized deฤŸerlendirme ile baลŸlamak
  2. gerรงek baฤŸlam uzunluklarฤฑyla benchmark almak
  3. vLLM veya SGLang gibi รถzel serving stackโ€™leri kullanmak
  4. production รถlรงeฤŸi veya uzun baฤŸlam iรงin yรผksek bellekli altyapฤฑ ayฤฑrmak :contentReference[oaicite:22]{index=22}

Yazฤฑlฤฑm Gereksinimleri

ร–nerilen ortam:

  • Python 3.10+
  • Linux
  • CUDA destekli GPU ortamฤฑ
  • ลžu runtimeโ€™lardan biri:
    • Transformers
    • vLLM
    • SGLang
    • KTransformers

Resmi model kartฤฑ ayrฤฑca JetLLMLite-3.6 iรงin en gรผncel transformers sรผrรผmรผnรผn gerektiฤŸini ve multimodal kullanฤฑm iรงin torchvision ile pillow kurulu olmasฤฑ gerektiฤŸini belirtir. :contentReference[oaicite:23]{index=23}

Yaygฤฑn baฤŸฤฑmlฤฑlฤฑklar:

  • torch
  • transformers
  • torchvision
  • pillow

Hฤฑzlฤฑ BaลŸlangฤฑรง

Transformers kurulumu:

pip install "transformers[serving]"

Temel yรผkleme รถrneฤŸi:

from transformers import pipeline

pipe = pipeline(
    "image-text-to-text",
    model="Jetlink/JetLLMLite-3.6",
    trust_remote_code=True
)

Serving ร–rnekleri

vLLM

vllm serve Jetlink/JetLLMLite-3.6

SGLang

docker run --gpus all \
  --shm-size 32g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=<secret>" \
  --ipc=host \
  lmsysorg/sglang:latest \
  python3 -m sglang.launch_server \
      --model-path "Jetlink/JetLLMLite-3.6" \
      --host 0.0.0.0 \
      --port 30000

Docker Model Runner

docker model run hf.co/Jetlink/JetLLMLite-3.6

Uzun BaฤŸlam Notlarฤฑ

JetLLMLite-3.6 yerel olarak 262.144 token destekler ve 1.010.000 tokenโ€™a kadar geniลŸletilebilir. Resmi model kartฤฑ uzun baฤŸlam kullanฤฑmฤฑ iรงin YaRN tabanlฤฑ rope scaling yapฤฑlandฤฑrma rehberi de iรงerir. :contentReference[oaicite:24]{index=24}

Gรผรงlรผ Yรถnler

  • gรผรงlรผ kodlama ve agentic kabiliyetler
  • geliลŸmiลŸ frontend ve repository-level reasoning
  • multimodal destek
  • รงok uzun yerel baฤŸlam
  • geniลŸletilebilir ultra uzun baฤŸlam
  • modern MoE mimarisi
  • popรผler aรงฤฑk inference frameworkโ€™leriyle uyumluluk :contentReference[oaicite:25]{index=25}

Sฤฑnฤฑrlamalar

  • deployment tipine gรถre altyapฤฑ ihtiyacฤฑ ciddi olabilir
  • uzun baฤŸlam inference bellek baskฤฑsฤฑnฤฑ bรผyรผk รถlรงรผde artฤฑrabilir
  • multimodal iลŸ yรผkleri ek kaynak tรผketir
  • local kullanฤฑmฤฑn pratikliฤŸi quantization ve runtime seรงimine รงok baฤŸlฤฑdฤฑr
  • gerรงek dรผnya latency ve throughput deฤŸerleri framework ve donanฤฑm yapฤฑlandฤฑrmasฤฑna gรถre ciddi biรงimde deฤŸiลŸir

Kapsam DฤฑลŸฤฑ / Dikkat Gerektiren Kullanฤฑmlar

ร‡ฤฑktฤฑlar ลŸu alanlarda insan kontrolรผ olmadan kullanฤฑlmamalฤฑdฤฑr:

  • tฤฑbbi karar verme
  • hukuki tavsiye
  • gรผvenlik kritik otomasyon
  • yรผksek riskli finansal kararlar
  • doฤŸrulama olmadan tam otonom aksiyonlar
  • korumasฤฑz hassas production iลŸ akฤฑลŸlarฤฑ

ฤฐnsan incelemesi, tool doฤŸrulamasฤฑ ve politika kontrolleri gรผรงlรผ ลŸekilde รถnerilir.

Lisans

Bu depo, upstream sรผrรผmle aynฤฑ lisansฤฑ takip eder.

  • Lisans: Apache-2.0

Modeli yeniden daฤŸฤฑtฤฑyor, fine-tune ediyor, quantize ediyor veya baลŸka ลŸekilde deฤŸiลŸtiriyorsan; kullanฤฑmฤฑnฤฑn upstream lisans ve attribution gereklilikleriyle uyumlu olduฤŸundan emin olmalฤฑsฤฑn. :contentReference[oaicite:26]{index=26}

Atฤฑf

Orijinal upstream model:

  • Qwen/Qwen3.6-35B-A3B

Bu depo, kurum tarafฤฑndan yรถnetilen bir kopyadฤฑr ve orijinal upstream kaynak deฤŸildir.

Atฤฑf / Citation

Bu modeli araลŸtฤฑrma, deฤŸerlendirme veya production dokรผmantasyonunda kullanฤฑyorsan, lรผtfen orijinal Qwen3.6 sรผrรผmรผne atฤฑf yap.

Feragatname

Bu depo, kurumsal kullanฤฑm amacฤฑyla paketleme, isimlendirme veya daฤŸฤฑtฤฑm odaklฤฑ bazฤฑ deฤŸiลŸiklikler iรงerebilir.

Resmi gรผncellemeler, benchmark detaylarฤฑ, uzun baฤŸlam ayarlarฤฑ ve upstream sรผrรผm notlarฤฑ iรงin orijinal Qwen model kartฤฑna bakฤฑlmalฤฑdฤฑr. :contentReference[oaicite:27]{index=27}

Downloads last month
14
Safetensors
Model size
36B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Jetlink/JetLLMLite-v1.1-36B-A3B

Finetuned
(440)
this model
Quantizations
1 model