AI Model Timeline

Open-weight multimodal MoE that Qwen ships as an early preview of the architecture behind Qwen4 — the same role Qwen3-Next played for Qwen3.5. A 125B main model plus a separate 51B N-gram embedding table, with 6B parameters activated per token. Four changes carry the release: a GDN + QSA hybrid attention stack (Gated DeltaNet compresses the history, Qwen Sparse Attention picks important context at micro-block granularity), a Gated Residual stream widened to four dynamically gated branches, the N-gram embedding table itself (looked up from local context, offloadable to host memory with asynchronous prefetching), and a Muon optimizer with a scaling law refitted for the new architecture. Native context is 262,144 tokens. Qwen reports roughly 1/9 the training cost of Qwen3.7-Plus. Because only 6B parameters are active and the embedding table can sit in system RAM, Unsloth's smallest quant runs on a 78 GB RAM / unified-memory machine without a discrete GPU. A production Qwen3.8-Flash for the QwenCloud API was announced at $0.16 / $0.47 per MTok but is not live yet, so no cloud entry is tracked; the license file ships with the weights and is not stated in the GitHub release.

Mid-size member of the Ornith-1.5 family: 35B total / 3B active MoE under MIT, trained with the same self-improvement loop as the 397B flagship and shipped with GGUF, MLX (including a 4-bit build) and NVFP4 checkpoints, so a quantised build fits a 24 GB consumer GPU. Vendor-reported at launch: Terminal-Bench 2.1 67.8 (Harbor/Terminus-2, 128K context, averaged over 5 runs) and SWE-bench Verified 79.0 (OpenHands harness, 256K context). Like the rest of the family it is a post-trained derivative of open weights, not an in-house pretraining run; the MIT grant covers Ornith's own weights, the upstream base terms govern the lineage.

Dense 27B natively multimodal model under Apache 2.0 — the companion release to the Qwen3.8 open weights and the one that actually fits a single consumer GPU. Takes text, images and video through a 27-layer vision encoder, with a native 262,144-token context extensible to ~1M via YaRN and switchable thinking. Qwen positions it for agentic work: the launch table shows the largest gains over Qwen3.6-27B in agentic coding, computer use and vision-language tasks. At NVFP4 the weights are about 15.7 GB, a Q4_K_M GGUF roughly 16.4 GB, so 24 GB of VRAM is the comfortable target.

General-availability build of DeepSeek-V4-Pro. Identical 1.6T total / 49B active MoE architecture and size as the April preview — re-post-trained, with the gains concentrated in agentic and software-engineering tasks. 1M-token context, MIT license.

NVIDIA's current open-weight release: a 30B / 3B-active hybrid model interleaving Mamba-2, MoE and select attention layers, tuned for throughput in agent loops rather than peak scores. Text-only with a 1M-token context, shipped under OpenMDW-1.1 together with training data and post-training recipes, in BF16, NVFP4 and community GGUF builds. A 4-bit quant runs on a single 24 GB consumer GPU. NVIDIA reports the largest gains over Nemotron 3 Nano on agentic evaluations; Artificial Analysis measures roughly 670 output tokens/s on pre-release infrastructure.

On-device agentic model from Liquid AI, 2.69B parameters with a 131K-token context window and tool calling, trained to work inside real agent harnesses rather than to top knowledge benchmarks. Liquid measures 220 tok/s on an Apple M5 Max, 113 tok/s on an AMD Ryzen AI Max+ 395 and roughly 30 tok/s on a phone, all under 2.5GB of memory, and reports it beating the 4x larger Qwen3.5-9B on ToolSandbox (77.83 vs 76.44) and on every instruction-following benchmark. No MMLU or GPQA figures have been published, so the benchmark fields stay null.

2.8T MoE (~50B active, 16 of 896 experts) native multimodal flagship with a 1M-token context window. Open weights released on Hugging Face 2026-07-27 under the bespoke Kimi K3 License — the largest open-weight model at release.

Tencent's Hunyuan 3 flagship: a 295B-parameter MoE activating 21B per token (192 routed experts with top-8 routing plus an always-active shared expert), 256K context window and three selectable reasoning-effort modes. Full release under Apache 2.0 — unlike the April preview license, without territorial restrictions.

Largest model of the Nemotron 3 family: 550B total / 55B active hybrid Mamba-Transformer MoE. Released with weights, training data and recipes under OpenMDW-1.1.

218B sparse Mixture-of-Experts (~25B active) open-weight flagship under Apache 2.0, runnable on as few as 2x H100 GPUs. Enterprise- and RAG-focused with strong multilingual coverage. Available as open weights and via Cohere's API. Early independent benchmarks (Artificial Analysis): GPQA Diamond ~76, HLE ~11, MMMU-Pro 63, Intelligence Index ~37 — flagged as estimates, AA's full evaluation is still ongoing.

Open-weights flagship from Xiaomi. 1.02T total / 42B active hybrid MoE with interleaved Sliding Window + Global Attention (6:1, 128-token window). 1M context, MIT-licensed for commercial use and fine-tuning without additional authorization. Day-0 SGLang/vLLM support. Reports SWE-bench Pro 57.2 and ClawEval 64% Pass³ at ~70K tokens per trajectory.

Preview release, superseded by DeepSeek-V4-Flash-0731 on 2026-07-31. Smaller 284B total / 13B active MoE sibling of V4-Pro, sharing hybrid attention and hyper-connection architecture.

Open vision-language-action model for humanoid robots, released in early access under Apache 2.0 — the first fully commercially licensed model of the GR00T line. A vision-language backbone is paired with a diffusion transformer head that denoises continuous actions via flow matching. NVIDIA credits the generalization and language-following gains over N1.6 to 20,000 hours of EgoScale human egocentric video in pretraining. At 3B parameters it is small enough to run on local hardware, which is why it also appears in the Local tab.

Open-weight agentic MoE: 230B total / 10B active, 256 experts, ~200K context. Coding and tool-use focus.

Major architectural upgrade over Qwen3. 397B MoE (17B active) flagship with 1M context, natively multimodal, 8-19x higher decoding throughput. Covers 200+ languages.

Trillion-parameter reasoning model with adaptive tool use and test-time scaling. Highest-capability open model from Alibaba.

Open multilingual model from the EU-funded EuroLLM / UTTER consortium, covering 35 languages including all 24 official EU languages. Dense decoder-only transformer (22B; a 9B variant released earlier) pretrained on 4T+ tokens with balanced European-language data. Positioned as a sovereign European base model for translation and multilingual understanding. Base and instruct variants on Hugging Face.

Successor to DeepSeek V3. 671B MoE (37B active) with DeepSeek Sparse Attention for long-context efficiency and large-scale agentic tool-use pipeline.

Experimental model pushing the boundaries of open weights. Achieved strong reasoning and coding scores rivaling closed models.

Incremental update with better instruction following and improved benchmark scores across the board.

Code-specialized open model optimized for software engineering. Achieved top-tier SWE-bench scores among open-weight models.

Mistral's first reasoning model, released as open weights at 24B under Apache 2.0. The larger Magistral Medium sibling announced the same day is API-only and has no published weights — it is tracked as a separate cloud entry.

235B MoE flagship with hybrid thinking modes.

Open MoE multimodal model with 128 experts. Frontier-level quality at efficient inference cost with 17B active parameters.

111B Dense Enterprise-Flagship (command-a-03-2025), 256k Kontext, 23 Sprachen, Fokus auf Tool-Use, agentische Workflows und Translation. Läuft auf wenigen GPUs, offene Gewichte. Direkter Vorgänger von Command A+. Vectara HHEM 9.3% Halluzination.

First open reasoning model using reinforcement learning. Matched OpenAI o1 on math and coding reasoning tasks without supervised fine-tuning for chain-of-thought.

Efficient 70B text model matching Llama 3.1 405B on key benchmarks.

Suite of editing tools including Fill, Depth, Canny, and Redux.

Compact 2.5B parameter variant optimized for consumer GPUs. Runs on hardware with 4-16 GB VRAM.

Flagship 8.1B parameter open-weight image model with improved quality and prompt adherence. Available in standard and Turbo variants.

Flagship 72B model rivaling GPT-4o and Claude 3.5 on key benchmarks.

Open-weight text-to-image model from Black Forest Labs. Guidance-distilled 12B param model for non-commercial use.

123B parameter open model with strong coding capabilities. Competitive with GPT-4o and Claude 3.5 Sonnet on code generation tasks.

Massive 405B open-source model competing with frontier closed models.

Next-gen architecture with Multimodal Diffusion Transformer.

Major upgrade with 72B flagship, competitive with leading Western models.

104B Dense, RAG- und Tool-Use-optimiertes Open-Weight-Flagship (CC-BY-NC), 128k Kontext, citation-fähiges Retrieval. Starkes Grounding beim Zusammenfassen (Vectara HHEM 6.9% Halluzination für den 08-2024-Build), akademische Benchmarks dagegen schwach. Direkter Vorläufer der Command-A-Linie.

314B MoE (86B active) open-weight model released under Apache 2.0. Largest open-weight model at the time of release.

Multi-size release from 0.5B to 110B with strong multilingual support.

First mainstream open Mixture-of-Experts model. Used 8 expert networks with only 2 active per token, achieving 70B-quality at 13B inference cost.

Efficiency breakthrough from Mistral AI. A 7B model that outperformed Llama 2 13B on all benchmarks using Grouped-Query Attention and Sliding Window Attention.

Alibaba's first open-source large language model.

First commercially licensed open LLM. Made open-weight models viable for business use and kickstarted the open-source LLM ecosystem.

Revolutionary open-source text-to-image model that democratized AI art.