The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestQwen · Alibaba3.8
Qwen · AlibabaText / LLM

Qwen releases 2.4T-parameter open MoE with 95B active

Alibaba's Qwen team pushes its largest sparse model yet, activating 95 billion parameters per token from a 2.4-trillion-parameter pool.

Aug 8, 2026
Major releaseOther
Qwen3.8-2.4T-A95B

Qwen, the AI group inside Alibaba, has published Qwen3.8-2.4T-A95B on Hugging Face — a text and reasoning model that stands as the team's largest open Mixture-of-Experts release to date. The headline figures are in the name: roughly 2.4 trillion total parameters, with about 95 billion active for any given token.

That sparse design is the whole point. Rather than firing every parameter on every forward pass, an MoE routes each token to a small subset of specialized "experts." The result is a model with the knowledge capacity of a very large network but the per-token compute closer to a mid-sized dense model. In practice, the 95B active count is what determines inference cost, while the 2.4T total sets the ceiling on what the model can store.

Why it matters

  • Scale meets openness. Frontier-scale MoE models have largely stayed behind closed APIs; shipping weights at this size continues Qwen's push to keep serious capability in the open ecosystem.
  • Reasoning focus. The model is tagged for both general text and reasoning, signaling it's aimed at the harder problem-solving workloads that increasingly define the top tier.
  • Deployment realism. A 2.4T-parameter checkpoint is not something most teams will run casually; expect adoption to concentrate among well-resourced labs and inference providers.

One caveat worth flagging: the release carries an "other" license rather than a standard permissive one, so teams should read the terms on the model page before building on it. As always with a fresh drop, independent benchmarks and real-world testing will determine how the raw parameter counts translate into everyday performance.

Sources

  • Qwen/Qwen3.8-2.4T-A95B

    Hugging Face

    Visit
  • Qwen3.8 Weights Released

    Hacker News

    Visit
OlderIBM's Granite 4.2 30B adds reasoning and tool callsIBM · Text / LLM · 3 months agoNewerdots3-note preview brings audio and vision to one modelDots Studio · Any-to-Any · 3 months ago

Get the model

Hugging FaceHacker News

Specs

Parameters2400B · MoE
Active params95B active
Context window262K tokens
Size4.9 TB
PrecisionBF16
ArchitectureQwen3_5MoeForCausalLM
LicenseOTHER
Downloads29.8K
Likes1.4K

Can you run it?

Datacenter only — about 1.5 TB even at 4-bit.

  • BF16 (as published)

    Multi-node cluster

    4.9 TB
  • 8-bit

    Multi-node cluster

    2.6 TB
  • 4-bit

    Multi-node cluster

    1.5 TB
Your machine
GGUF builds
  • Mixture-of-experts: all 2450B parameters must be resident even though only 95B are active per token — active count affects speed, not memory.
  • KV cache sized for 8,192 tokens of context; longer prompts need proportionally more.

Based on Hugging Face weights metadata. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

Text / LLMReasoning

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More from Qwen · Alibaba

All Qwen · Alibaba releases →
Qwen · Alibaba/Speech → Text

Qwen-Audio-3.1-ASR Adds Speaker-Aware Transcription

Alibaba's new open audio-language model transcribes and diarizes speech across 30 languages under an Apache 2.0 license.

Oct 19, 2026
Qwen-Image-2.1-Turbo
Qwen · Alibaba/Text → ImageRuns on a laptop

Qwen-Image-2.1-Turbo targets faster generation

Alibaba's Qwen team ships a distilled, few-step variant of its image model built for speed.

Oct 9, 2026
Qwen-Image-2.1
Qwen · Alibaba/Text → ImageRuns on a laptop

Qwen-Image-2.1 Adds RGBA to Image Generation

Alibaba's Qwen team updates its open image model with transparency support and joint text-to-image and editing capabilities.

Sep 14, 2026

More in Text / LLM

All Text / LLM →
Aion-1.0-Instruct-2.3B
Microsoft/Text / LLMRuns anywhere

Microsoft's Aion 1.0 Instruct is a compact 2.3B chat model

A small, multilingual instruction-tuned model aimed at chat and summarization lands on Hugging Face.

Nov 18, 2026
SenseNova-Loop-9B
SenseTime/Vision-LanguageRuns on a laptop

SenseTime Releases Looped 9B Multimodal Model

The company's new SenseNova-Loop-9B pairs vision-language capabilities with a recurrent looped-transformer design.

Nov 17, 2026
Hunyuan Hy4-flash
Tencent/Vision-LanguageMulti-GPU server

Tencent releases Hunyuan Hy4-flash multimodal MoE

The lightweight, mixture-of-experts model adds vision and text instruction-following to Tencent's open Hunyuan lineup.

Nov 17, 2026