Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Janus family

Janus-1.3B

Hugging Face ↗ Follow changes

deepseek-ai MIT (deepseek) out 2 years ago

2.09B parameters, several kinds of input and output.

Verified 2026-10-25 Rank #677 in LLMs · score 72.5 method v1 Sources 3 Downloads 30d 1.1k · #592 of 619 tracked

Can I run it?1.3GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?MIT (deepseek)Permissivelicence file2026-10-24
Do people use it?1.1kdownloads, 30 days#592 of 619 we track · below the median+3 likes, downloads −28% this week610 likesHugging Face2026-10-24

Worth reading

Picked by rule from 6 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 106 HN pts

Show HN: Janus – Go binary that runs GGUF models via Vulkan on AMD/Intel/Nvidia

Janus is a single Go binary that runs .gguf models on your machine and exposes an OpenAI-compatible API.

  • The build script pins llama.cpp to version b11146 because internal struct layouts match that specific tag.
  • On Windows, Go cannot overwrite a running executable, requiring the process to be stopped in Task Manager before rebuilding.
github.com 2026-10-01 106 HN pts · 19 comments 774 words about the Janus family family
Latest · 2 years ago

Janus: Decoupling visual encoding for multimodal understanding and generation

DeepSeek released Janus-Pro on January 27, 2025, as an advanced version of Janus that improves multimodal understanding and visual generation.

  • A tokenizer configuration bug previously prevented classifier-free guidance from functioning properly, degrading visual generation quality.
  • The generation pipeline requires setting a classifier-free guidance weight of 5 to achieve the intended image synthesis results.
github.com 2024-10-20 36 HN pts · 3 comments 2.1k words about the Janus family family

6 best of 6 tracked

The 2 picks above, plus these 4. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeDeepSeek releases Janus Pro, a text-to-image generator [pdf]github.com · 2025-01-27
  2. DiscussionJanus Pro 1B running 100% locally in-browser on WebGPUold.reddit.com · 2025-01-28
  3. ArticleProject Janus: Now Open Sourcetechcommunity.microsoft.com · 2025-03-26 · 4 HN pts
  4. ArticleShow HN: DeepSeek Janus Pro 7B APIeachlabs.ai · 2025-01-28

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 1 quantized builds and 21 Spaces.

  • As a service0no API provider listed
  • On own hardware1quantized builds
  • Built on33 fine-tunes · 0 adapters · 0 merges
  • In apps21Hugging Face Spaces

Top Spaces running it: reach-vb/2024-ai-timeline ♥ 467 · deepseek-ai/Janus-1.3B ♥ 162 · blanchon/JanusPro ♥ 26 · Deadmon/janusAI ♥ 7

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 1.3 GB at INT4, 2.5 GB at FP8, 5.0 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.3 GB of weights (estimate; KV cache and activations extra)INT41.3 GBFP8: about 2.5 GB of weights (estimate; KV cache and activations extra)FP82.5 GBBF16: about 5.0 GB of weights (estimate; KV cache and activations extra)BF165.0 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.3 GB of weights (estimate; KV cache and activations extra)INT41.3 GBFP8: about 2.5 GB of weights (estimate; KV cache and activations extra)FP82.5 GBBF16: about 5.0 GB of weights (estimate; KV cache and activations extra)BF165.0 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

3 dated articles and 8 repo commits so far

Teal is the weights (2024-10-18), amber triangles are 3 dated articles (each links to it), grey ticks are repo commits.

NovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-01-27: 1 repo commit2024-11-14: 1 repo commit2024-10-20: 3 repo commits2024-10-18: 3 repo commitsShow HN: Janus – Go binary that runs GGUF models via Vulkan on AMD/Intel/Nvidia · 2026-10-01Janus: Decoupling visual encoding for multimodal understanding and generation · 2024-10-20Project Janus: Now Open Source · 2025-03-26Weights 2024-10-18 · 2024-10-18Weights 2024-10-18now
NovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-01-27: 1 repo commit2024-11-14: 1 repo commit2024-10-20: 3 repo commits2024-10-18: 3 repo commitsShow HN: Janus – Go binary that runs GGUF models via Vulkan on AMD/Intel/Nvidia · 2026-10-01Janus: Decoupling visual encoding for multimodal understanding and generation · 2024-10-20Project Janus: Now Open Source · 2025-03-26Weights 2024-10-18 · 2024-10-18Weights 2024-10-18now

What people built on it

4 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

4 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

1 quantized (Hugging Face model tree)3 fine-tunes (Hugging Face model tree)1 quantized3 fine-tunes
1 quantized (Hugging Face model tree)3 fine-tunes (Hugging Face model tree)1 quantized3 fine-tunes

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

1 cited sentence

Summary generated, cited

Janus-1.3B is a 2.09B parameter multimodal model designed for unified understanding and generation.[1][2]

  1. parameters human
  2. “unified understanding and generation”

Generated 2026-10-07 15:15 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

2 use cases

What it is good at 2 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

17 fields

Listed in LLMs

Facts

Pipeline tagany-to-any
Librarytransformers
Model typemulti_modality
Dtypebfloat16
Params by dtypeBF16: 2,089,297,547
Size bandM
MoEno
PapersarXiv:2410.13848
LicenseMIT (deepseek) (mit)
Hub snapshot2026-10-25

Servability

SGLang

Servability

SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Janus-1.3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)5.0 GB · Single GPU (<24 GB)
FP8 (1 byte/param)2.5 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.3 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

4 variants

Lineage

Original model Family: deepseek-ai/Janus 4 derivatives on the Hub

quantized: 1finetune: 3

Top variants (4 shown, ranked by downloads)

Relation
4 rows
Variant Relation Quant Params Band License Author Likes Downloads
onnx-community/Janus-1.3B-ONNX quantized — — — other onnx-community 16 4.5k
Georgio007/SE-BIM-01 finetune — — — apache-2.0 Georgio007 1 0
BroAmbula/Curriculum_Prediction finetune — — — apache-2.0 BroAmbula 0 30
calebbenchoff/Have_Hope finetune — — — — calebbenchoff 0 0

Repository changelog

8 commits
changed in the last 7 daysNew on the watch list: 2026-10-23all changes

Repository changelog 8 commits

DateCommitSHA
2025-01-27 Update processor_config.json 56c70f4
2024-11-14 Add image normalization in preprocessor. 98b20d0
2024-10-20 Update README.md 2cf2f40
2024-10-20 Upload tokenizer_config.json b238799
2024-10-20 Update special_tokens_map.json 5294d91
2024-10-18 Update README.md (#1) 825ec51
2024-10-18 first commit 08b9a5d
2024-10-18 initial commit 247e00a

Articles and coverage

6 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

Project Janus: Now Open Source
techcommunity.microsoft.com2025-03-26authorityabout the Janus familyHN 4 pts · 1 comments8,243 words
Microsoft open-sourced Project Janus on Feb 27, 2025, to optimize O-RAN networks with dynamic service models and real-time telemetry. The release includes the jbpf library and jrt-controller, enabling control at microsecond time scales and access to fine-grained L1 telemetry. Previously announced at MWC 2024, the framework supports anomaly detection, interference mitigation, and fast vRAN power savings.written, reviewed, cited
  • Project Janus enables adding new functionality to operational RAN deployments without causing service disruption.
  • The framework allows developers to exert control over RAN behavior at microsecond time scales.
  • Project Janus allows real-time dynamic access to L1 telemetry, letting the telemetry type be controlled based on future AI use cases.

Analysis2

Show HN: Janus – Go binary that runs GGUF models via Vulkan on AMD/Intel/Nvidia
github.com2026-10-01authorityHN front pageabout the Janus familyHN 106 pts · 19 comments774 words
Janus is a single Go binary that runs .gguf models on your machine and exposes an OpenAI-compatible API. It requires Go 1.25+ and supports local inference via Vulkan or CPU fallback without requiring Python, Docker, or Ollama. The build script pins llama.cpp b11146, and the server listens on port 8990 under an MIT license.written, reviewed, cited
  • The build script pins llama.cpp to version b11146 because internal struct layouts match that specific tag.
  • On Windows, Go cannot overwrite a running executable, requiring the process to be stopped in Task Manager before rebuilding.
  • The first reply often takes 10–60 seconds because the model is loading into VRAM during startup.
Janus: Decoupling visual encoding for multimodal understanding and generation
github.com2024-10-20authorityabout the Janus familyHN 36 pts · 3 comments2,112 words
DeepSeek released Janus-Pro on January 27, 2025, as an advanced version of Janus that improves multimodal understanding and visual generation. The model incorporates optimized training strategies, expanded training data, and scaling to larger model sizes to achieve significant advancements. JanusFlow, released on November 13, 2024, introduces a minimalist architecture integrating autoregressive language models with rectified flow for image generation. All models support a sequence length of 4096 and are available on Hugging Face.written, reviewed, cited
  • A tokenizer configuration bug previously prevented classifier-free guidance from functioning properly, degrading visual generation quality.
  • The generation pipeline requires setting a classifier-free guidance weight of 5 to achieve the intended image synthesis results.
  • Engineers must set the image token count per image to 576 and the image size to 384 pixels during the generation process.
All coverage tracked for this model (6) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Show HN: Janus – Go binary that runs GGUF models via Vulkan on AMD/Intel/Nvidiagithub.com2026-10-012026-10-07hn
Project Janus: Now Open Sourcetechcommunity.microsoft.com2025-03-262026-10-07hn
Show HN: DeepSeek Janus Pro 7B APIeachlabs.ai2025-01-282026-10-07hn
Janus Pro 1B running 100% locally in-browser on WebGPUold.reddit.com2025-01-282026-10-07hn
DeepSeek releases Janus Pro, a text-to-image generator [pdf]github.com2025-01-272026-10-07hn
Janus: Decoupling visual encoding for multimodal understanding and generationgithub.com2024-10-202026-10-07hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • gemma-4-E2B-it · 5.12B
    same size band (M), same task (any-to-any), permissive license
  • gemma-4-E2B · 5.12B
    same size band (M), same task (any-to-any), permissive license
  • Qwen2.5-Omni-3B · 5.54B
    same size band (M), same task (any-to-any)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:15 UTC · Last verified 2026-10-25

Follow Janus-1.3B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.