Model release
DeepSeekMoE · text + image
DeepSeek-V4.1-Flash
The September model that changes the deployment diagram
Large multimodal services investigating conditional memory and distinct prefill/decode paths.
552Breported total
The open model reference / September 2026
Understand open models.
Choose what to run.
You do not need to know what “35B-A3B” means yet. Start with the basics, then compare model releases, useful tasks and the hardware they need.
Explore by research lab
See the lineage, not just the latest name.
Compare all lab profiles. Coverage is the curated set in this edition, not each lab’s entire historical catalog.
New to open weights?
The download is one part of the system.
01Architecture
The layers and connections that define the computation.
02Weights
The learned tensor values, sometimes converted or quantized.
03Tokenizer + runtime
Software that formats inputs, loads the model and generates outputs.
04Your service
Hardware, request state, tools, access controls and evaluation.
You can run an open-weight model locally, operate a server, or call a provider that hosts it. Downloadable weights do not automatically include the materials needed to reproduce training, and local inference does not prevent an agent’s tools from sending data elsewhere.
Release timeline
Newest first. Useful older baselines included.
Compare up to three checkpoints. A later date tells you where a model sits in the timeline; your workload decides whether it is the better choice.
39 of 39 checkpoints. Dates identify the sourced release or announcement event. Month-only records stay in their month; their position does not imply a day.
September 2026
August 2026
Release note
Z.aiMoE · text + image + video
GLM-5.3-Flash
Smaller than GLM-5.3, still a multi-GPU model
Visual coding agents and long-context multimodal services with compatible kernels.
320Breported total18B active
Open-weight release
Alibaba QwenMoE · text + image + video
Qwen3.8-Flash-Next
The memory placement is part of the architecture
Multimodal services that can test GPU versus host placement and newer hybrid kernels.
180Breported total6B active
Announcement
Z.aiMoE · text
GLM-5.3
A post-training release with a changed adoption decision
Controlled coding-agent and authorized defensive-security evaluations on a cluster.
743Breported total39B active
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.8-27B
A current dense baseline before you scale out
Mid-sized multimodal services and controlled comparisons with Qwen3.6-27B.
27Breported total
General availability
DeepSeekMoE · text
DeepSeek-V4-Pro-0813
Keep the production update distinct from the preview
Reproducing the August Pro behavior and comparing it with the preview under one harness.
Not verifiedreported total
Open-weight release
Alibaba QwenMoE · text
Qwen3.8-2.4T-A95B
Large-scale text reasoning, with large-scale residency
Cluster-scale text reasoning and coding evaluation with a carefully pinned checkpoint.
2,400Breported total95B active
Model-card month
MetaDense · text + image
Muse-Glimmer-30B
A local agent needs more than a small download
Private on-device agents using text and image inputs on a capable workstation.
29.6Breported total
July 2026
Update release
DeepSeekMoE · text
DeepSeek-V4-Flash-0731
A post-training update you can compare fairly
Upgrades from V4 Flash preview and reproducible same-architecture comparisons.
284Breported total13B active
Open-weight release
Moonshot AIMoE · text + image + video
Kimi-K3
A cluster model with a new attention design
Long repository tasks and multimodal agent evaluations on a high-memory cluster.
2,800Breported total104B active
June 2026
Release note
Z.aiMoE · text
GLM-5.2
A million-token setting backed by index sharing
Large text-agent services investigating long-context retrieval and coding.
743Breported total39B active
Open-weight release
Moonshot AIMoE · text + image + video
Kimi-K2.7-Code
A coding upgrade without a new size class
Repository editing, test-driven repair and long agent runs.
1,000Breported total32B active
Research release
Google DeepMindDiffusion MoE · text + image + video
diffusiongemma-26B-A4B-it
A different generation loop, not a drop-in speed setting
Experiments comparing completion time, refinement work and answer quality.
25.2Breported total3.8B active
Model release
Google DeepMindDense · text + image + video + audio
gemma-4-12B-it
One model path for text, vision and audio
Applications combining text, image, video and audio inputs.
11.95Breported total
Announcement
MiniMaxMoE · text + image + video
MiniMax-M3
Sparse reads do not automatically mean a small cache
Long-context coding and multimodal agent evaluation with supported MSA kernels.
428Breported total23B active
April 2026
Preview release
DeepSeekMoE · text
DeepSeek-V4-Flash
The smaller V4 preview that made cache design central
Architecture study and controlled comparisons against the later Flash releases.
284Breported total13B active
Preview release
DeepSeekMoE · text
DeepSeek-V4-Pro
Trillion-parameter storage with sparse per-token work
Large-cluster text evaluations and study of expert-parallel serving.
1,600Breported total49B active
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.6-27B
A clean dense-generation comparison
Comparative multimodal evaluation under one serving configuration.
27Breported total
Model release
Moonshot AIMoE · text + image + video
Kimi-K2.6
Keep the hardware fixed; compare the completed work
Existing K2.5 deployments evaluating agent and multimodal task improvements.
1,000Breported total32B active
Open-weight release
Alibaba QwenMoE · text + image + video
Qwen3.6-35B-A3B
A smaller expert store for coding-agent experiments
Multimodal coding-agent prototypes and upgrades from Qwen3.5-35B-A3B.
35Breported total3B active
Release note
Z.aiMoE · text
GLM-5.1
Longer agent runs need process-level evaluation
Long coding and engineering workflows with recorded tool traces and acceptance tests.
744Breported total40B active
Open-weight release
Google DeepMindMoE · text + image + video
gemma-4-26B-A4B-it
Sparse compute in a workstation-sized family
Mid-sized text, image and video deployments and dense-versus-MoE studies.
25.2Breported total3.8B active
Open-weight release
Google DeepMindDense · text + image + video
gemma-4-31B-it
The dense control for Gemma’s MoE experiment
Dense multimodal serving and comparisons that keep the model family fixed.
30.7Breported total
Open-weight release
Google DeepMindDense · text + image + video + audio
gemma-4-E2B-it
“Effective” is a compute label, not the download size
Small-device text, vision and audio experiments with a supported Gemma runtime.
5.1Breported total
Open-weight release
Google DeepMindDense · text + image + video + audio
gemma-4-E4B-it
More effective capacity, with tables still to place
Multimodal edge evaluation where E2B is too limited and the larger footprint is acceptable.
8Breported total
March 2026
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.5-0.8B
A cheap baseline for narrow tasks
Constrained text and visual tasks with task-specific acceptance tests.
0.8Breported total
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.5-2B
A small step up when the tiniest model misses too much
Constrained text and visual tasks with task-specific acceptance tests.
2Breported total
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.5-4B
A practical small-model starting point
Constrained text and visual tasks with task-specific acceptance tests.
4Breported total
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.5-9B
Test whether the middle size earns its memory
Constrained text and visual tasks with task-specific acceptance tests.
9Breported total
February 2026
Open-weight release
Alibaba QwenMoE · text + image + video
Qwen3.5-122B-A10B
A middle step before the largest Qwen3.5 model
Multi-GPU multimodal evaluation where the 35B model misses important cases.
122Breported total10B active
Open-weight release
Alibaba QwenDense · text + image + video
Qwen3.5-27B
Dense compute alongside a sparse sibling
Controlled dense-versus-MoE tests with text and visual inputs.
27Breported total
Open-weight release
Alibaba QwenMoE · text + image + video
Qwen3.5-35B-A3B
The easiest MoE memory lesson to make concrete
Learning MoE deployment and testing a comparatively modest multimodal service.
35Breported total3B active
Open-weight release
Alibaba QwenMoE · text + image + video
Qwen3.5-397B-A17B
The launch model that established Qwen3.5’s hybrid design
Large-model baselines and study of the family’s initial architecture.
397Breported total17B active
Release note
Z.aiMoE · text
GLM-5
A large sparse-attention foundation for the GLM line
Large text-agent baselines and reproducing early GLM-5 results.
744Breported total40B active
Announcement
Alibaba QwenMoE · text
Qwen3-Coder-Next
Executable feedback matters as much as model size
Code tools that can run tests, inspect repository state and measure completed changes.
80Breported total3B active
January 2026
Model release
Moonshot AIMoE · text + image + video
Kimi-K2.5
The multimodal starting point for the K2 series
Reproducing K2-era evaluations and understanding large multimodal MoE deployment.
1,000Breported total32B active
Release note
Z.aiMoE · text
GLM-4.7-Flash
A compact way to learn latent-attention serving
Smaller text coding experiments and study of latent-attention caches.
30Breported total3B active
August 2025
Open-weight release
OpenAIMoE · text
gpt-oss-120b
A single large-memory GPU is a useful deployment class
Self-hosted text reasoning and controlled studies of scale versus the 20b model.
117Breported total5.1B active
Open-weight release
OpenAIMoE · text
gpt-oss-20b
A compact reasoning baseline that still earns its place
Local text reasoning, tool-use prototypes and experiments with reasoning effort.
21Breported total3.6B active
What counts as evidence here?
Architecture facts link to pinned model cards and configurations. Deployment rows name the artifact, precision, hardware count and runtime conditions. We label publisher memory claims separately from engine-maintainer recipes.
Our model assessments are editorial judgments, not an independent quality leaderboard. No GPU benchmark was run. Exact release days are not invented when only a month or announcement is documented. Download the public JSON records →