Models/K2 Horizon/K2 Horizon MoVA 36B A4B

Text / IFM

K2 Horizon MoVA 36B A4B

K2-Horizon-MoVA-36B-A4B is the sparse member of the K2-Horizon family, a Mixture-of-Experts model with Mixture-of-Values attention (MoVA). It stores 36B parameters and runs 4B parameters per token. The final checkpoint has been released; intermediate checkpoints, training data, recipe, and training code are stated as forthcoming. It supports text generation, is licensed under Apache-2.0, and has a native 524,288-token context from midtraining stages onward. The model card claims frontier-class results at 4B active parameters on agentic and reasoning benchmarks, and competitive performance against closed frontier models, but these claims are descriptions from the source rather than independently verified here.

EstimatedRepository opened Sep 1, 2026Source checked 9/23/2026Version: e9f33cb4

LOCALRENTED GPUOPEN WEIGHTS

Freshness 84 out of 100. Repository age: 44 days old, Sep 1, 2026. Measures age, not quality.

Popularity 59 out of 100. 15,925 downloads in 30 days; 416 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

At a glance

Parameters: 37.44B
Architecture: k2_horizon
Context length: 524,288
License: apache-2.0
Disk space: 69.7 GB
Software: vLLM, Transformers

View the model source

MY HARDWARE

Will it run on your machine?

Save your machine to see a personalized rating and its reasoning.

Add my hardware

Best for

Text generation tasks
Agentic workflows
Reasoning tasks
Long-context processing up to 524,288 tokens
Studying capability changes across intermediate checkpoints, once released

Tradeoffs

Only the final checkpoint has been released; intermediate checkpoints are not yet available.
Training data, recipe, and training code are stated as forthcoming but not yet released.
The model is described as supporting English; other languages are not specified.
Performance claims are from the official model card and should be independently verified.
No numerical hardware requirements or compatibility details are provided in the source.

Ways to run it

vLLM · Linux

Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.

vllm serve IFM/K2-Horizon-MoVA-36B-A4B  --revision main  --tensor-parallel-size 2  --enable-expert-parallel  --trust-remote-code  --dtype bfloat16  --reasoning-parser k2_horizon  --tool-call-parser k2_horizon  --enable-auto-tool-choice

Official instructions

Transformers · See official guide

Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "IFM/K2-Horizon-MoVA-36B-A4B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", dtype="bfloat16", low_cpu_mem_usage=True, trust_remote_code=True
)

inputs = tokenizer("Explain why long-context evaluation is difficult.", return_tensors="pt").to(model.device)
inputs.pop("token_type_ids", None)
outputs = model.generate(**inputs, max_new_tokens=32768, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Official instructions

Explore its uses

SAFETENSORSTEXT-GENERATION

Keep exploring

Text

Prism ML

Bonsai 2 · 27B

A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.

Freshness 90 out of 100. Release age: 28 days old, Sep 17, 2026. Measures age, not quality.

MODEL SIZE27.36B

LOCALRENTED GPU

Released Sep 17, 2026

Text

DeepSeek

DeepSeek V4.1 Flash

A multimodal reasoning model with a compressed key-value cache, published as open weights.

Freshness 87 out of 100. Repository age: 35 days old, Sep 10, 2026. Measures age, not quality.

Arena Open Source Text: rank 5 of 225. Ranked model: deepseek-v4.1-flash-max. Export dated Oct 13, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.+3

LOCALRENTED GPU

Repository opened Sep 10, 2026

Text

Qwen

Qwen3.8

Explore 4 versions of Qwen3.8. Compare model sizes, packages, hardware requirements, and available setups.

4 versionsCompare sizes & formats

LOCALRENTED GPU

Compare versions →