The Open-Weights Bubble Race: 76 Models, Three Years, One Frontier

In 2006, Hans Rosling stood on a TED stage and turned global health data into a sport. Countries became bubbles — income on one axis, life expectancy on the other — racing across decades in minutes. You could see China rise and HIV push back an entire continent.

This post does the same thing to open-source AI. Models are the bubbles. They enter at release and race through capability space from 2023 to today: 76 open-weights releases plus 15 closed flagships for context, all cross-checked against vendor model cards in October 2026.

Press Play above to follow the race, or drag the timeline to explore a date.

How to read it#

• X → MMLU % (general knowledge). Y → Intelligence Index (composite overall score, switchable to Arena Elo, HumanEval, or context window).
• Bubble size → parameters (MoE active count in the tooltip). Color → organization.
• The dotted gold trail is the open-source frontier: every model that set a new all-time high. It never includes closed models — they are context, not contenders.
• The default lens shows the trailing 12 months with auto-zoom. Older bubbles slide out left as the camera pushes into the frontier cluster. Untick both boxes for the full cumulative view.
• The 🔒 commercial button overlays GPT, Claude, and Gemini as dashed bubbles. Their params are undisclosed, so their sizes are estimates.

Three eras in ninety seconds#

2023 — the Cambrian explosion. LLaMA leaks the recipe in February and the field detonates: Alpaca, Vicuna, Falcon, then Llama 2 70B takes the first real crown in July. Mistral 7B lands in September and proves small can win — 60.1% MMLU from 7.3B params. Mixtral closes the year by making MoE mainstream. Watch the closed overlay here: GPT-4 sits alone, far ahead of everything open, for most of the year.

2024 — consolidation. Llama 3 70B, Qwen2 72B, then Llama 3.1 405B in July — the first open model that could look GPT-4-class hardware in the eye. Context windows jump from 8K to 128K as the default. Small models stop being toys: Phi-3-mini and Gemma 2 27B hold respectable positions at a fraction of the size.

2025–2026 — scale and convergence. DeepSeek-R1 (January 2025) resets everyone's assumptions about reasoning. Kimi K2, GLM-4.5, and Qwen3 push the open frontier past 1400 Arena Elo — territory that belonged only to closed labs a year earlier. Then the giants arrive: 975B-parameter Inkling, the 2.4T-parameter Qwen3.8, NVIDIA's Nemotron 3 family all the way to Super VL. By late 2026 the open frontier and the closed frontier are the same neighborhood.

The frontier, end to end#

Every model that moved the open maximum, in order: LLaMA 13B → Vicuna 13B → Falcon 40B → Llama 2 70B → Mistral 7B → Mixtral 8×7B → Qwen 72B → Mixtral 8×22B → Llama 3 70B → Qwen2 72B → Llama 3.1 70B → Llama 3.1 405B → DeepSeek-V3 → DeepSeek-R1 → DeepSeek V3.2-Exp → Nemotron 3 Ultra → GLM-5.2 → Qwen3.8 2.4T-A95B.

Two things stand out. First, the crown changes labs constantly — Meta, Mistral, Alibaba, DeepSeek, NVIDIA, Zhipu all hold it at some point. No moat survived contact with the next release. Second, the gaps between successive frontiers shrink over time: the 2026 jumps are refinements, not revolutions.

The full dataset#

Every number behind the bubbles, filterable by license. Same file the chart renders from — no hidden columns.

Methodology and honest caveats#

MMLU and HumanEval come from official model cards and papers (instruct figures where published, else base). Later vendors stopped reporting HumanEval in favor of SWE-Bench and LiveCodeBench, which are mapped approximately — treat post-2025 coding numbers as directional.

Arena Elo is an LMArena snapshot near launch; the live board drifts daily. The chart's Intelligence Index is a 0–100 animation composite, not the official Artificial Analysis index (different scale and versioning) — use it for relative motion, not absolute claims. Closed-model parameter counts are undisclosed, so those bubble sizes are estimates. Qwen3-Max is excluded: closed API-only, not open weights.

References#

• LLaMA: Touvron et al. “LLaMA: Open and Efficient Foundation Language Models.” arXiv 2302.13971
• Llama 2: Touvron et al. “Llama 2: Open Foundation and Fine-Tuned Chat Models.” arXiv 2307.09288
• Gemma: Gemma Team. “Gemma: Open Models Based on Gemini Research and Technology.” arXiv 2403.08295
• Meta Llama model cards: Llama 3 / 3.1 / 3.2 / 3.3. github.com/meta-llama/llama-models
• Mistral AI newsroom: Mistral 7B, Mixtral, Large 2, Ministral, Devstral, Small 3.1. mistral.ai/news
• Qwen: Qwen1–Qwen3.8 blogs and technical reports. qwen.ai/blog
• DeepSeek: DeepSeek-V2 / V3 / R1 / V4 papers. github.com/deepseek-ai/DeepSeek-V3
• Moonshot AI: Kimi K2, K2.5, K3. github.com/MoonshotAI/Kimi-K2
• Z.ai: GLM-4.5, 4.6, 5, 5.2. huggingface.co/zai-org
• OpenAI: gpt-oss model card; GPT-4 technical report. openai.com/index/gpt-oss-model-card · arXiv 2303.08774
• Google: Gemma 2 / 3 / 4 releases. ai.google.dev/gemma/docs/releases
• Thinking Machines: Inkling, Inkling-Small. thinkingmachines.ai/news/introducing-inkling
• NVIDIA: Nemotron 3 Super; Nemotron 3.5 Super VL + Bolt. research.nvidia.com · blogs.nvidia.com
• Databricks: DBRX. databricks.com/blog
• AllenAI: OLMo 3. allenai.org/blog/olmo3
• LMArena leaderboard. lmarena.ai
• Artificial Analysis. artificialanalysis.ai