Open weights are a deployment decision, not a magic label

The native text/image-to-text checkpoint is published under MIT, with a 552B backbone, 196B Engram conditional memory and a native one-million-token context label. September 10 is the hosted launch date; the exact native checkpoint publication day is unverified.

Licence

MIT

Scale

552B MoE backbone + 196B Engram conditional memory; 8B active prefill / 16B decode

Practical fit: Check actual model files, mixed precision, conditional-memory placement, cache and the chosen backend. The backbone count alone omits Engram storage; a generic uniform-Q4 calculation does not establish native pack size or desktop fit. The 8B/16B active counts are not total weight-memory requirements.

Deployment: The repo now publishes a minimal inference reference with vision, Engram, MoE and DSpark paths; generation is plain autoregressive sampling. Its self-test uses uninitialized weights and checks shapes/kernel plumbing, not real model behaviour. Native effort is 1–100 with low=50/high=75/max=100 and default high=75; this does not establish the hosted API default or output cap. Production serving and an actual local install were not tested.

Xiaomi's September open-weight release is a 1.02T MoE with a 1M-token context. The named RL checkpoint supplies the architecture/licence facts; separate Xiaomi API docs verify the hosted model ID and completion budget.

Licence

MIT

Scale

1.02T total / 42B active MoE

Practical fit: A hosted endpoint may have separate prices and terms from the checkpoint; verify the exact provider and account terms before budgeting or redistribution.

Deployment: Weights are published as the RL checkpoint (MiMo-V2.6-Pro-RL). The first-party API ID is mimo-v2.6-pro, with a 131,072-token completion budget including reasoning. Self-hosting the downloadable checkpoint requires datacentre-scale serving.

Mistral Medium 3.5 is a 128B dense open-weight model. Its model card documents a 256K context; the hosted API and weight licence have separate terms.

Licence

Modified MIT

Scale

128B dense

Practical fit: A dense open-weight option for teams evaluating a single-checkpoint deployment; performance, serving cost and licence fit depend on the chosen setup.

Deployment: Dense 128B needs multi-GPU memory at useful precision; review the licence modifications before commercial redistribution.

Moonshot describes Kimi K3 as a 2.8T-parameter open-weight multimodal model with a 1M-token context. Its checkpoint is governed by the Kimi K3 License, not a permissive MIT grant.

Licence

Kimi K3 License (custom terms)

Scale

2.8T total MoE; active-parameter count not stated here

Practical fit: A datacentre-scale checkpoint for teams evaluating provider independence and long-context work, not casual local inference.

Deployment: The checkpoint is roughly a datacentre deployment; read the exact [Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE) before use or redistribution. Active-parameter count is not stated here.

DeepSeek's July API version documentation says the `deepseek-v4-flash` alias was updated to V4-Flash-0731. The API alias was later retired on September 10 and temporarily routes to V4.1-Flash; the available first-party receipt does not establish the downloadable checkpoint's modality or architecture.

Licence

Unverified in available first-party evidence

Scale

Unverified in available first-party evidence

Practical fit: Historical API/checkpoint identity only. For a current hosted route, consult DeepSeek's live model and pricing pages; do not assume a local checkpoint is equivalent.

Deployment: The available first-party receipts do not establish exact checkpoint parameter count, license, or a supported self-hosting recipe.

An open multimodal mixture-of-experts release designed for broad agent and reasoning experiments.

Licence

Apache 2.0

Scale

975B total / 41B active MoE

Practical fit: Useful for teams building an open-model evaluation lane, especially where modality breadth matters.

Deployment: Total parameter count makes quantisation, serving, and memory planning central to the decision.

A smaller Inkling checkpoint aimed at bringing much of the family’s capability into a more practical serving envelope.

Licence

Apache 2.0

Scale

276B total / 12B active MoE

Practical fit: The better first experiment for teams that want to test the family without starting with the full checkpoint.

Deployment: Smaller does not mean laptop-sized; estimate memory and throughput from the actual release artefacts.

A large sovereign-model release from LG AI Research for teams evaluating Korean and enterprise-focused open infrastructure.

Licence

Apache 2.0

Scale

750B foundation model

Practical fit: Interesting for regional language, sovereignty, and research use cases where a broad open ecosystem matters.

Deployment: Confirm the actual checkpoint, serving recipes, and hardware requirements before treating it as a production option.

A smaller policy-adaptive safety model designed to sit around a generation system rather than replace it.

Licence

Apache 2.0

Scale

3B classifier

Practical fit: Useful for moderation, routing, and safety layers where a compact model can reduce latency and cost.

Deployment: Evaluate false positives and policy fit on your own content; a classifier is not a complete safety programme.

A compact on-device agent and tool-calling model for latency-sensitive local workflows.

Licence

LFM Open License v1.0

Scale

2.6B

Practical fit: A better candidate for edge and private inference than the frontier-sized releases in this list.

Deployment: The LFM Open License v1.0 contains revenue-based commercial-use terms; read the exact license and evaluate tool-call reliability at the target memory and quantization level.

An open vision-language-action model aimed at embodied and autonomous systems rather than general chat.

Licence

OpenMDW-1.1

Scale

34B total / 32B VLM backbone + action expert

Practical fit: Relevant to robotics and physical-world evaluation, not a drop-in chatbot replacement.

Deployment: Hardware, simulator, sensor, and safety validation are part of the product decision.

Meta describes Muse Glimmer as an open-weight model distilled from Muse Spark for local agent workflows and says it can run on a single consumer GPU; actual fit depends on the checkpoint, quantization and runtime.

Licence

Apache 2.0

Scale

~29.6B dense (incl. ~1.8B perception encoder)

Practical fit: The practical local-agent path when Muse Spark API is too heavy or closed for your deployment.

Deployment: Check the exact available artifact and runtime before sizing hardware; validate multimodal and agent behavior on your workflow.

InclusionAI’s cost-focused Flash MoE tier for high-frequency hybrid reasoning and agent loops.

Licence

MIT

Scale

124B total / 5.1B active MoE

Practical fit: A budget open-weight option when you need MoE efficiency and can host the expert set.

Deployment: The exact checkpoint config has 262,144 positions; the vendor describes extension to 1M. Thinking is enabled by default and can be disabled per request. The platform launched the native API July 23; July 24 dates the article, while the exact public weights release day is unverified. Current hosted prices and limits depend on the selected route.

Official Hugging Face checkpoint for Qwen3.8-2.4T-A95B: 2.4T total / 95B active parameters, native 262K context and text input. Its hosted Qwen3.8-Max counterpart has a separate service contract and capabilities.

Licence

Qwen3.8-Max License

Scale

2.4T total / 95B active MoE

Practical fit: The self-host path for Qwen3.8 flagship weights. It is not a drop-in for the hosted Qwen3.8-Max API's vision, tools, or 1M default context.

Deployment: This is a datacentre-scale checkpoint. Its card documents YaRN extension toward 1,010,000 tokens. The exact [Qwen3.8-Max License](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE) sets separate commercial-service and scale terms; review it before deployment.

Apache 2.0 Qwen3.8 dense VLM with native 262K context and image/video input — the practical local sibling of hosted Qwen3.8-Max.

Licence

Apache 2.0

Scale

27B dense

Practical fit: The first Qwen3.8 local target when 2.4T-A95B is too large and you can accept a dense 27B serving envelope.

Deployment: Current Alibaba pricing lists the hosted qwen3.8-27b endpoint; the earlier coming-soon model-card sentence is not current availability. Native context is 262,144 and a hosted 1M contract is separate. Vendor DeepSWE/SWE-bench Pro results used Claude Code and are not public-board results.

DeepSeek's V4-Pro-0813 release has separate open-weight and hosted API surfaces. The official API lists peak/off-peak billing; do not transfer API terms to the weights.

Licence

MIT

Scale

1.6T total / 49B active (MoE)

Practical fit: An evaluation candidate when testing V4-Pro against V4.1-Flash on harder reasoning or coding. DeepSeek's V4.1 comparison is vendor-reported, not a matched independent ranking. Pin the 2026-08-13 V4-Pro-0813 checkpoint when comparing it with the April preview.

Deployment: The model card reports a 1.6T/49B-active MoE. Hosted API id is `deepseek-v4-pro`; the live price page separately lists peak $1.32/$3.96 and off-peak $0.66/$1.98 per million input/output tokens.

Z.ai's open-weight GLM-5.2 checkpoint is documented with a 1M-token context; its API profile lists a 128K max output and $1.40/$4.40 token rates. Parameter count remains unpublished or conflicting in available primary materials.

Licence

MIT

Scale

Not published on catalog row

Practical fit: Useful for open-weight / token-list deployments that can carry serving and integration work. GLM Coding Plan traffic for 5.2/5.1 now routes to GLM 5.3.

Deployment: The $1.40/$4.40 rates apply to the hosted token API, not to self-hosting. Size stays unverified — do not plan hardware from an unpublished or conflicting parameter count.

Ant’s open-weight multimodal Flash variant adds image/video understanding and tool workflows to a sparse 124B model.

Licence

MIT (publisher metadata)

Scale

124B total / 5.5B active (MoE)

Practical fit: A multimodal self-hosting option when the full 124B weight set fits your serving hardware.

Deployment: Exact native config has 131,072 positions; the publisher’s 262,144 recipe enables YaRN factor 2. Hosted API launch dates do not prove first public weights dates. MIT is declared in repository metadata; the requested raw licence file is missing. Provider APIs and route limits are separate from self-hosting.

Ant’s 7.9B open-weight text MoE for local or edge reasoning, code assistance and tool use, with 1.3B active parameters.

Licence

MIT (publisher metadata)

Scale

7.9B total / 1.3B active (MoE)

Practical fit: A smaller local text and tool-use checkpoint; budget for the full 7.9B weights rather than 1.3B activation.

Deployment: Exact native config has 131,072 positions; the publisher’s 262,144 recipe enables YaRN factor 2. Hosted API launch dates do not prove first public weights dates. MIT is declared in repository metadata; the requested raw licence file is missing. Provider APIs and route limits are separate from self-hosting.