KAISTA CLOUD / PRIVATE BETAOpen Documentation ↗
Kaista CloudAI COMPUTE CLOUD
Start building
EXPLORE / MODEL DIRECTORY

Explore models.Know what runs.

Review 49 open-weight and coding models curated by Awesome Free Models. A listing is a research reference—not a promise that the model is callable through Kaista Cloud.

49Open-model referencesThird-party catalog · CC0
OPEN-MODEL REFERENCE LIBRARY

open weight

32 results · Page 1
open-weight
Meta Superintelligence LabOPEN REFERENCE

Muse Glimmer (Meta Superintelligence Lab)

REFERENCE · awesome-free-models/muse-glimmer

Aug 2026. 30B dense multimodal model distilled from Muse Spark for local agentic workflows. Multi-step reasoning, tool use, vision, and failure recovery on consumer hardware (4-bit quant fits 24GB VRAM). 131K context. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
MetaOPEN REFERENCE

Llama 4 Scout / Maverick

REFERENCE · awesome-free-models/llama-4-scout-maverick

Meta's latest MoE generation. Scout: 109B, 10M context. Maverick: 402B, 1M context. Native multimodal.

open-weight
Request model review →Hugging Face ↗
open-weight
DeepSeekOPEN REFERENCE

DeepSeek V4 Pro

REFERENCE · awesome-free-models/deepseek-v4-pro

Apr 2026. 1.6T MoE (49B active). SWE-bench Verified 80.6% (top open-weight). 1M context. MIT license.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
DeepSeekOPEN REFERENCE

DeepSeek V4

REFERENCE · awesome-free-models/deepseek-v4

Core generation with extreme cost-efficiency. 1M context. MIT license.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
DeepSeekOPEN REFERENCE

DeepSeek-V4-Flash

REFERENCE · awesome-free-models/deepseek-v4-flash

Apr 2026. Efficiency-focused variant. 284B total (13B active). 1M context. MIT license.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
GoogleOPEN REFERENCE

Gemma 4 31B / 26B MoE / E4B / E2B

REFERENCE · awesome-free-models/gemma-4-31b-26b-moe-e4b-e2b

Fully permissive Apache 2.0. 256K context, native multimodal. New standard for open-weight.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
Thinking Machines LabOPEN REFERENCE

Inkling (Thinking Machines Lab)

REFERENCE · awesome-free-models/inkling

Jul 2026. 975B MoE (41B active). Leading US open-weight model. Native multimodal (text, image, audio). Apache 2.0. 1M context.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
Zhipu AIOPEN REFERENCE

GLM-5.2 (Zhipu AI)

REFERENCE · awesome-free-models/glm-5-2

744B MoE model optimized for autonomous coding and engineering tasks. 1M-token context. MIT license.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
ByteDanceOPEN REFERENCE

LongCat-2.0 (ByteDance)

REFERENCE · awesome-free-models/longcat-2-0

Large-scale open-weight model for heavy agentic coding. MIT license.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
MiniMaxOPEN REFERENCE

MiniMax M3

REFERENCE · awesome-free-models/minimax-m3

Frontier-tier 1M context, native multimodal + computer use. MSA architecture.

open-weight
Request model review →Hugging Face ↗
open-weight
StepFunOPEN REFERENCE

Step 3.7 Flash (StepFun)

REFERENCE · awesome-free-models/step-3-7-flash

May 2026. Apache 2.0. Native multimodal (image+video), strong agentic performance. Efficient enough for high-end local hardware.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
Moonshot AIOPEN REFERENCE

Kimi K3 (Moonshot AI)

REFERENCE · awesome-free-models/kimi-k3

Jul 2026. 2.8T-parameter MoE (896 experts, ~50B active). World's largest open-weight model. 1M context, native vision + video. #1 Frontend Code Arena. Modified MIT license. Weights released Jul 27.

open-weightModified MIT
Request model review →Hugging Face ↗
open-weight
Moonshot AIOPEN REFERENCE

Kimi K2.6 (Moonshot AI)

REFERENCE · awesome-free-models/kimi-k2-6

Apr 2026. 1T-parameter MoE model. Modified MIT license. Exceptional coding (SWE-Bench ~54%) and multi-agent swarm orchestration.

open-weightModified MIT
Request model review →Hugging Face ↗
open-weight
AlibabaOPEN REFERENCE

Qwen 3.6-35B-A3B

REFERENCE · awesome-free-models/qwen-3-6-35b-a3b

Apr 2026. MoE variant with only 3B active parameters. Extremely efficient for consumer hardware. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
Shanghai AI LabOPEN REFERENCE

InternLM 3 (Shanghai AI Lab)

REFERENCE · awesome-free-models/internlm-3

Early 2026. Strong long-context reasoning and agentic performance. Competitive in open-weight benchmarks.

open-weight
Request model review →Hugging Face ↗
open-weight
XiaomiOPEN REFERENCE

MiMo-V2.5-Pro (Xiaomi)

REFERENCE · awesome-free-models/mimo-v2-5-pro

Apr 2026. 1.02T-parameter MoE (42B active). Optimized for complex agentic tasks, coding, and long-context.

open-weight
Request model review →Hugging Face ↗
open-weight
Moonshot AIOPEN REFERENCE

Kimi K2.7 Code (Moonshot AI)

REFERENCE · awesome-free-models/kimi-k2-7-code

Jun 2026. 1T MoE specialized for long-running coding agents. +21.8% over K2.6 on coding benchmarks. Modified MIT.

open-weightModified MIT
Request model review →Hugging Face ↗
open-weight
NVIDIAOPEN REFERENCE

Nemotron 3 Super (NVIDIA)

REFERENCE · awesome-free-models/nemotron-3-super

May 2026. 120B total (12B active). 1M context. Published weights, data, recipes, and eval infra. NVIDIA Open Model License.

open-weightNVIDIA Open Model License
Request model review →Hugging Face ↗
open-weight
MicrosoftOPEN REFERENCE

Phi-4 14B (Microsoft)

REFERENCE · awesome-free-models/phi-4-14b

2025. Compact 14B dense model. Strong reasoning and code. MIT license. Excellent for on-device and small deployments.

open-weightMIT license
Request model review →Hugging Face ↗
open-weight
PrismMLOPEN REFERENCE

Bonsai 8B (PrismML)

REFERENCE · awesome-free-models/bonsai-8b

Apr 2026. Groundbreaking 1-bit quantized model. Extremely efficient for edge and consumer hardware (Apple Silicon).

open-weight
Request model review →Hugging Face ↗
open-weight
VIDRAFTOPEN REFERENCE

Aether-7B-5Attn (VIDRAFT)

REFERENCE · awesome-free-models/aether-7b-5attn

Jul 2026. 100% open foundation model (weights, data, code, logs). 7B MoE (~3B active) with heterogeneous attention. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
MistralOPEN REFERENCE

Mistral Large 3 (Mistral)

REFERENCE · awesome-free-models/mistral-large-3

Jun 2026. 675B MoE (41B active). European multilingual flagship. Frontier-class reasoning, native multimodal. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
MistralOPEN REFERENCE

Mistral Small 3.1 (Mistral)

REFERENCE · awesome-free-models/mistral-small-3-1

Mar 2025. Versatile 24B multimodal model. Strong text performance with native image understanding and 128K context. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
MistralOPEN REFERENCE

Mistral Small 4 (Mistral)

REFERENCE · awesome-free-models/mistral-small-4

Mar 2026. Hybrid MoE (6.5B active params) unifying instruction, reasoning, and multimodal capabilities. Efficient frontier-class model. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
CohereOPEN REFERENCE

Command A+ (Cohere)

REFERENCE · awesome-free-models/command-a

May 2026. Enterprise multimodal MoE optimized for sovereignty and multilingual RAG across 48 languages. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
ETH Zurich / EPFLOPEN REFERENCE

Apertus 1.5 (ETH Zurich / EPFL)

REFERENCE · awesome-free-models/apertus-1-5

Jul 2026. Fully open LLM (weights, data, training code). 8B and 70B with image understanding, thinking mode, and tool use. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗
open-weight
TencentOPEN REFERENCE

Hy3 (Tencent)

REFERENCE · awesome-free-models/hy3

Jul 2026. 295B MoE (21B active). Strong reasoning and agentic performance. Competes with models 2-5x its size. Apache 2.0.

open-weightApache 2.0
Request model review →Hugging Face ↗