Zen · Open models

Open models.
Yours to run.

Zen is the open-weight family behind Hanzo: seven generations on one line, from a laptop to the datacenter. Download the weights, run them privately, or call the same family through Hanzo Cloud.

Run Zen 6 Download weights

Explore the lineage

5.8

Now

Available nowZen 627.3B dense. 262K context. Text, images and video.Run Zen 6

Research previewZen 7The next generation, open to researchers by request.Request access

Available now

Zen 6.
Open now.

A 27.3 billion parameter dense model. Hybrid attention, a long context, images and video, and a speculative drafter bundled with the weights. The interesting part is not that it is the biggest. It is that you can run it.

Agentic coding

Long repositories, tool use, long context, and fast speculative decoding.

Multimodal work

Zen 6 reads text, images, and video in one context.

27.3BParameters

262KNative context

1MWith YaRN

On one DGX Spark, the weight card records 62.4 tokens per second on code alone and 141.2 with the drafter. That figure is published with the weights. It is not a Hanzo harness.

Zen 6 Flash zenlm/zen6

Zen 6 Flash

27B.
Laptop-sized.

The same generation, packed ternary. 26,895,998,464 language-model weights in a 5.95 GB file, 1.77 bits per weight, 9.0 times smaller than the 53.79 GB FP16 representation. Text and images. Not video.

FP16 representation53.79 GB

Ternary1.77 bits / weight

5.95 GBDense packed

1.77Bits per weight

262KNative context

Apple Silicon or one GPU, through llama.cpp and Metal. The card records 47.2 tokens per second alone and 92.4 with the drafter on M4 and M5 Max, in 7.84 GB with the projector and a 32K cache. Published with the weights. Not a Hanzo harness.

zenlm/zen6-flash

Measured by Hanzo

A frontier model.
On your desk.

Hanzo ran Zen 6 on three machines it owns. These are the readings, in tokens a second.

NVIDIA Grace-Blackwell

121 GB unified · NVFP4 weights

~2,400tok/s prefill

~43tok/s decode

50.6 tok/s decoding code · 112 tok/s across 8 requests

AMD Strix Halo

128 GB unified · 4-bit build

1,150–1,400tok/s prefill

~34tok/s decode

262K context, in two slots

Apple silicon

unified memory · MLX build

500–690tok/s prefill

64–81tok/s decode

Hanzo's runs, 23 September 2026. Decode is one request unless a line says otherwise, and each machine runs its own build of Zen 6. Where the Hanzo engine's prefill on CUDA beats llama.cpp is on the inference benchmark.

Engine prefill on CUDA

Open weights

APIs can disappear.
Weights don't.

Download them, read them, fine-tune them, and run them where you want. Zen 6 is Apache-2.0, published by Zoo Labs Foundation, a 501(c)(3) non-profit.

Zen weightsYour laptopYour GPUHanzo Cloud

The family on Hugging Face Source

2023 to now

Seven generations. One line.

Foundation, then context, then a family, then scale, then the models that actually shipped. Each generation inherits the last, and the current one is the practical one.

Zen 1

Q4 2023

HistoricalFoundationDense multilingual experiments at 7B, 14B and 72B, to prove an open foundation could scale. These are not models the gateway serves.One dense block.

Zen 2

Q2 2024

HistoricalGo longer128K context, the first production deployments, and code and math specialists. From research model to production model. The context got longer. The architecture did not get more decorative.The same block, wider.

Zen 3

2024

Open weightsMore than languageThe line branched: sparse experts, vision, audio, image, retrieval and safety. This is where Zen becomes a family. None of these ids are in the live catalog. Open-weight specialist artifacts are still published.One line, six branches.

LanguageVisionAudioImageRetrievalSafety

zenlm/zen3-omni zenlm/zen3-tts zenlm/zen3-guard

Zen 4

2025

RetiredScale upFrontier-scale experts and a longer context. Retired. Not a current production model, and not a public download.Large, and sparse.

Zen 5

ShippedAgentic generationCode, long context, and smaller and host-specific variants. Older pages described a much larger routed checkpoint. It was not published. These are the models that were.Paths for tools and hosts.

zen5 zen5-coder zen5-evo zen5-flash zen5-mini zen5-pro zen5-spark

5.8

Zen 5.8

Served onlyServing bridgeA served-only bridge between Zen 5 and Zen 6, with no public weights. The gateway does not list it now.A thin serving step.

Zen 6

CurrentDense againZen 6 is a practical dense model: hybrid attention, a long context, images and video, speculative decoding, and Zen 6 Flash, a ternary build that fits a laptop. Available now.Smaller, and solid.

zen6 zen6-flashThe current generation

Zen 7

Research previewResearch previewZen 7 has no public weights and no callable id. Its architecture, size, modalities and date are not stated. Research preview — not currently available.The release path, unresolved past the preview.

What comes next is still research.

Train

Attest

EvaluateNot published

ReleaseNot published

The path to a release is training, then attestation, then evaluation, then publication. Nothing past the preview is published.

Request research access The family on Hugging Face

Today

Every sense.
One family.

Zen MoDE, a Mixture of Diverse Experts: a model for each job, from generating and coding to seeing, speaking, retrieving and guarding. Grouped by job, from the models the gateway lists. Hosted means the id is in the catalog. Open weights means a public repository for that exact id.

Generate

zen6

CurrentOpen weightsHosted

API 1MWeights

zen6-flash

CurrentOpen weightsHosted

API 1MWeights

zen-free

Hosted

API 1M

zen5

Hosted

API 1M

zen5-evo

Hosted

API 1M

zen5-flash

Hosted

API 1M

zen5-mini

Hosted

API 1M

zen5-pro

Hosted

API 1M

zen5-spark

Hosted

API 1M

Code

zen5-coder

Hosted

API 1M

See

zen-vl

Hosted

API 1M

Speak and listen

zen-scribe

Hosted

zen-voice-mini

Hosted

Retrieve

zen-embedding

Hosted

API 33K

zen-embedding-2

Hosted

API 8K

zen-rerank

Hosted

API 33K

Guard

zen-guard

Hosted

API 128K

The full Zen catalogue

Two artifacts

Open weights, and a hosted id.

The same name can be two things. A hosted id is not documented as the same bytes as a public repository with a similar name.

Open weights

Download. Self-host. Modify.

Hosted

An API call. Managed. Metered.

Call it

One call.
Any SDK.

The gateway speaks the OpenAI API, so the SDK you already use works. Point it at api.hanzo.ai and ask for "zen6". The weights are the download.

python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.hanzo.ai/v1",
    api_key="hk-...",  # your Hanzo API key
)

resp = client.chat.completions.create(
    model="zen6",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

bash

curl https://api.hanzo.ai/v1/chat/completions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "zen6", "messages": [{"role": "user", "content": "Hello"}]}'

Try in HanzoDownload weights

Run it yourself

Three ways.

The commands are the ones on the current weight cards.

Download

hf download zenlm/zen6

hf download zenlm/zen6-flash

Local

Zen 6 Flash. llama.cpp and Metal. Apple Silicon or one GPU.

hf download zenlm/zen6-flash

The file names are the loader's.

bash

llama-server \
  -m Ternary-Bonsai-2-27B-PQ2_0.gguf \
  --mmproj Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf \
  -md Bonsai-2-27B-DFlash2-Q8_0.gguf \
  --spec-draft-n-max 3 \
  -c 32768 \
  --port 8080

Production

Zen 6. SGLang, with the bundled drafter.

bash

hf download zenlm/zen6 --local-dir zen6

python3 -m sglang.launch_server \
  --model-path ./zen6 \
  --speculative-draft-model-path ./zen6/dflash2 \
  --speculative-num-steps 3 \
  --speculative-algorithm DFLASH \
  --kv-cache-dtype fp8_e5m2 \
  --context-length 1048576 \
  --port 30000 \
  --host 0.0.0.0

Specifications

Zen 6 beside Zen 6 Flash.

Card facts only. A cell with no published value is left blank.

Zen 6Zen 6 Flash

Parameters

Zen 627.3 billion, dense, plus a vision encoder

Zen 6 Flash26,895,998,464 language-model weights

Weight format

Zen 6NVFP4 on the MLP and LM head, FP8 on attention

Zen 6 FlashTernary, 1.77 bits per weight, 5.95 GB packed

Native context

Zen 6262,144

Zen 6 Flash262,144

Extended context

Zen 61,048,576 with YaRN

Zen 6 Flash—

Text

Zen 6Yes

Zen 6 FlashYes

Images

Zen 6Yes

Zen 6 FlashYes

Video

Zen 6Yes

Zen 6 Flash—

Runs on

Zen 6A GPU, with SGLang. Measured on one DGX Spark.

Zen 6 FlashA laptop or one GPU, with llama.cpp and Metal

The API catalog lists 1M for zen6 and 1M for zen6-flash. The weight cards state 262,144 tokens native, and YaRN to 1,048,576 for Zen 6 only.

Layers, heads, and the files

Zen 6 has 64 layers, 48 of them linear attention and 16 full attention, one full layer every fourth. Hidden size 5,120. 24 query heads and 4 key and value heads, head size 256. Vocabulary 248,320. The speculative drafter ships in the weights.Zen 6 Flash also publishes a 7.21 GB file at 2.14 bits per weight, which is the one the kernels read, a 629 MB vision projector, and a 2.06 GB drafter. With the projector and a 32K cache it stays under 8 GB.

Scored

Scored where
it counts.

Hanzo scores the family on GPQA-Diamond, 198 graduate-level science questions. The publisher scores the open weights. The two are never mixed.

91.4%Zen 5.8 · GPQA-Diamond

90.8%Zen 5.8 Coder · GPQA-Diamond

98.2%Kept by Zen 6 Flash's ternary build

Zen 5.8 and Zen 5.8 Coder are Hanzo-scored, at $5.40 per million output tokens. The weights report 86.33 at full precision and 84.78 for the ternary build across 14 reasoning suites: the publisher's numbers, not a Hanzo harness. Zen 6 is not on the board yet, so this page does not borrow one.

Every benchmark

Your model should
survive your vendor.

Weights you can hold, runtimes you can choose, infrastructure you control.

DownloadInspectFine-tuneDeployFork

Zen reasons.

Open weights, yours to run. Routing and decisions are the other two families. They are compared on the models page.

Read

Models, benchmarks, research.

The weights you can run, where they are scored, and the writing.

Run Zen where you want.

Download the weights, run Zen privately, or use the same family through Hanzo.

Run Zen 6 Download weights

Zen 7 research preview

Open models.Yours to run.

Zen 6.Open now.

Agentic coding

Multimodal work

27B.Laptop-sized.

A frontier model.On your desk.

NVIDIA Grace-Blackwell

AMD Strix Halo

Apple silicon

APIs can disappear.Weights don't.

Seven generations. One line.

Zen 1

Zen 2

Zen 3

Zen 4

Zen 5

Zen 5.8

Zen 6

Zen 7

Every sense.One family.

Generate

Code

See

Speak and listen

Retrieve

Guard

Open weights, and a hosted id.

Open weights

Hosted

One call.Any SDK.

Three ways.

Download

Local

Production

Zen 6 beside Zen 6 Flash.

Scored whereit counts.

Your model shouldsurvive your vendor.

Zen reasons.

Models, benchmarks, research.

Models

Benchmarks

Research

Run Zen where you want.

Open models.
Yours to run.

Zen 6.
Open now.

27B.
Laptop-sized.

A frontier model.
On your desk.

APIs can disappear.
Weights don't.

Every sense.
One family.

One call.
Any SDK.

Scored where
it counts.

Your model should
survive your vendor.