Definition and categories
Open weights
The trained parameters are downloadable and runnable; the training data and code usually are not. The overwhelming majority of “open” models live here.
Llama (Meta) · Qwen (Alibaba) · DeepSeek · Mistral · gpt-oss (OpenAI)
Fully open
Weights plus training data, code and intermediate checkpoints: the model is reproducible end to end. Rare, and mostly research-led.
OLMo (Ai2) · Pythia (EleutherAI) · Nemotron (NVIDIA, close behind)
The licence axis
A separate question: from genuinely permissive terms to user caps, output restrictions and non-commercial clauses attached to otherwise-open weights.
Apache-2.0 · MIT · Llama Community · Gemma Terms · CC-BY-NC
What open means
“Open” is a combination of independent factors, not a single property. It's important to decide how reproducible and auditable the model is, what its commercial and legal terms allow, and whether you self-host or consume an API. A model can carry a permissive licence yet be impossible to reproduce, or ship with full data and code yet forbid commercial use.
The licences sort into four tiers. Permissive (Apache-2.0 / MIT) is now where the frontier lives: Qwen, DeepSeek, gpt-oss, GLM, Granite and ERNIE all ship under one or the other. Restrictive open weights attach conditions: Meta’s Llama Community Licence caps usage above 700 million monthly active users; xAI’s Grok-2 terms forbid using outputs to train other models. Non-commercial (Cohere’s Command and Aya, CC-BY-NC) is open to study but not to sell. Fully open (Ai2’s OLMo, with NVIDIA’s Nemotron close behind) publishes the data and recipes too. The centre of gravity has moved decisively toward the first tier: Meta’s licence, once the reference point for “open”, now reads as comparatively restrictive.
The case for — and the case against
For
Control. Data residency, privacy, compliance and air-gapped deployment: weights on your own hardware, in your own jurisdiction. A 2025 survey of ~1,500 IT leaders ranked data privacy the top AI-adoption barrier (53%).
Cost. No per-token markup at scale and the commodity tier is deflating. Equivalent-capability inference falls roughly 10× a year (a16z); Epoch AI measures a median ~50× across benchmarks.
Customisation. Full fine-tuning and weight surgery. For a specific task: code completion, transcription, embeddings, a single language. A fine-tuned open model is frequently the best cost-adjusted choice, closed models included.
Latency and offline. Sub-20ms local inference against 50–200ms cloud round-trips; edge and on-device operation with no connectivity requirement at all.
Vendor independence. No lock-in, no unilateral deprecation: a model you host cannot be switched off, repriced or retired from outside.
Against
The residual frontier gap. The hardest generalist reasoning and agentic-assistant work still favours the closed anchors. GPT-5, Claude Opus 4.5 (~81% SWE-bench Verified), Gemini 3 Pro, which benefit from massive reinforcement learning on real user feedback and proprietary agent infrastructure.
Operational burden. Self-hosting means GPUs, serving stacks, quantisation choices and on-call rotas. A GPU costs the same idle or busy: at low utilisation, effective cost can run ~10× the theoretical minimum.
Provenance and geopolitics. The strongest open models are disproportionately Chinese; government-device bans on DeepSeek span US federal agencies, Australia, Italy and Taiwan, and the regulatory environment around model origin is fluid.
The enterprise counter-signal. An a16z survey of enterprise CIOs (mid-2025) put open source at ~13% of workloads: down from ~19% six months earlier with OpenAI dominant in production.
Benchmark noise. Invalid-question rates of 2–42% across popular benchmarks, near-saturated maths tests and vendor best-case configurations.
Who builds the open frontier
The centre of gravity moved to China. The strongest open-weight models now come predominantly from Chinese labs: DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu/Z.ai’s GLM and MiniMax, with Baidu (ERNIE 4.5: a notable open pivot) and Tencent (Hunyuan). The single clearest structural change of 2025 was Qwen overtaking Llama as the most-used open family: 100,000+ fine-tuned derivatives on Hugging Face, more than Google and Meta combined, and roughly 700 million cumulative downloads by January 2026 (Xinhua).
Production traffic tells the same story. On OpenRouter, the combined token share of US providers fell from roughly 70% in mid-2025 to about 30% by mid-2026, with DeepSeek at times the single largest provider of token volume.
Meta pioneered the movement but appears to be retreating. Llama 4 (April 2025) fell behind the frontier it created, the flagship “Behemoth” never shipped, and in April 2026 Meta’s Superintelligence Labs released a proprietary model with no weights. Whether Llama 4 is Meta’s last open release is rumoured, not confirmed. Filling the gap, OpenAI re-entered open weights in August 2025 with gpt-oss (Apache-2.0). Its first open release since GPT-2. Google ships Gemma, Microsoft ships Phi, NVIDIA’s near-fully-open Nemotron line has become the strongest US open model, IBM targets the enterprise with Granite, and Ai2’s OLMo remains fully-open.
Outside the two poles: Mistral (FR) is Europe’s flagship, with permissive Small models plus code (Codestral, Devstral), reasoning (Magistral) and audio (Voxtral) lines; the UAE’s TII ships Falcon and the strongest Arabic models. Underneath it all sits scale: Hugging Face passed 2 million public models in 2025 and ~2.4–2.9M by mid-2026 with adoption extremely concentrated. The top 0.01% of models take roughly half of all downloads, and ~92% of downloads go to sub-1B-parameter models: embeddings, edge and pipeline use, not frontier chat.
The gap
Three independent trackers converge on the same conclusion: the gap is small, and no longer closing. Stanford HAI’s AI Index puts the top closed model 3.3% ahead on Chatbot Arena (up from 1.7% a year earlier, and 0.5% in mid-2024). Epoch AI measures the open-weights lag at 3–4 months. Artificial Analysis shows its composite intelligence gap collapsing from ~13 points in early 2025 to 3–6 points in 2026, with every one of its top-ranked open models coming from China-based labs.
Mathematics: effectively at parity. On competition maths (AIME 2025), the top open models — gpt-oss-120b at ~97.9% with tools, DeepSeek at ~96% — sit alongside or above GPT-5-class closed models. Agentic tool use: open has led. Kimi K2 Thinking posted state-of-the-art results on Humanity’s Last Exam (with tools) and BrowseComp, exceeding reported GPT-5 numbers, and stayed stable across 200–300 sequential tool calls. Coding: within roughly ten points. The best open models reach ~64–71% on SWE-bench Verified against the closed leader’s ~81% — behind, but still production-grade.
The single biggest driver of the 2025 leap was the open-sourcing of the reasoning paradigm. DeepSeek-R1 published the reinforcement-learning recipe for eliciting long chain-of-thought in January 2025; within a year, toggleable “thinking” modes were standard across Qwen, gpt-oss, Phi and Mistral’s Magistral. And because reasoning models distil into small dense ones (R1 shipped distillations from 1.5B to 70B), frontier-style reasoning now runs on consumer hardware.
A caution on the numbers. The AI Index found invalid-question rates of 2–42% across popular benchmarks; AIME is near-saturated, with memorisation concerns; contamination-resistant tests show drops of up to ~18 points against static ones; and vendors report best-case configurations (tools on, parallel sampling). Treat single-number leaderboard positions as indicative, not definitive.
Running it yourself
The tooling grew up. At the base sits llama.cpp and its GGUF format, a way to pack a whole model into a single file that runs on ordinary CPUs and consumer GPUs. Ollama and LM Studio are the desktop apps most people use to run a model on their own machine; vLLM and SGLang are what production servers run on. Hugging Face retired its own serving software, Text Generation Inference (TGI), in March 2026. Every layer speaks the OpenAI-compatible API, so the same application runs on a laptop or in a datacentre with a one-line change.
Mixture-of-experts made big models fit small hardware. The trick is that a model can carry an enormous total size but only fire a fraction of it per token, so memory stays modest while keeping quality high: a 671B-total model computes like a 37B one. That is why trillion-parameter open models now run on a single unified-memory workstation or high-memory server. The bottleneck has shifted off the models and onto the hardware.
Per token, open is dramatically cheap. Serving a Llama-70B-class model runs ~$0.10–1.04 per million tokens depending on provider, a ~10× spread that rewards shopping around, and DeepSeek-V3-class ~$0.20–0.56, against closed frontier APIs 10–50× higher. Prices for equivalent capability fall ~10× a year. In a mid-2026 price war, DeepSeek made a 75% cut to its flagship pricing.
Build-versus-buy depends on how utilised the hardware is, which has largely converged for everyday enterprise work. Rent per-token APIs when traffic is bursty or hard to predict. Host your own when volume is high and steady, which for a 70B-class model starts paying off somewhere in the hundreds of millions to low billions of tokens a month. Most organisations land on hybrid: a self-hosted baseline for the predictable load, with serverless to soak up the spikes. Teams already running their own inference juggle ~7 models on average, matching each job to the cheapest model that clears the bar.
Five findings
01China leads the open frontier.▶
Every top-ranked open model on Artificial Analysis comes from a China-based lab; Qwen displaced Llama as the ecosystem’s base layer; DeepSeek has at times been the largest single source of token volume on OpenRouter. The counterweights are regulatory and strategic. Government-device bans and procurement limits; as of July 2026 the US was considering, but had not enacted, restrictions on using Chinese open models. The July 2025 America’s AI Action Plan frames “leading open models founded on American values” as an explicit goal.
02The licence centre of gravity is now Apache-2.0 and MIT.▶
Qwen, DeepSeek, GLM, gpt-oss, Granite and ERNIE all ship genuinely permissive. Meta’s Llama Community Licence, the original reference for “open”, now reads as the restrictive outlier, and Google’s Gemma is reportedly (not confirmed) moving to Apache-2.0 with Gemma 4. For adopters this collapses most of the legal ambiguity that dogged early open deployment.
03Mixture-of-Experts made frontier-scale models locally viable.▶
Total parameters drive memory; active parameters drive speed. DeepSeek-V3’s 671B total computes through 37B per token; gpt-oss-120b activates 5.1B and runs on a single 80GB GPU.
04Local inference went mainstream.▶
Ollama at ~52 million monthly downloads; llama.cpp past 100,000 GitHub stars; ~135,000 GGUF files on Hugging Face; NPU phones decoding at >100 tokens per second. Production serving consolidated to vLLM and SGLang, with hardened air-gap builds from Red Hat and NVIDIA. The privacy, compliance and offline arguments now have a mature stack behind them. Reported enterprise switches include Coinbase, Airbnb, Uber and Shopify.
05Economics, not capability, is the battleground▶
Inference prices fall ~10× a year while a memory and power supply crunch inflates the hardware underneath. Capability has largely converged for common enterprise tasks; what remains is governance, provenance and total cost of ownership. The rational posture for nearly every organisation is neither “all open” nor “all closed”: closed frontier models for the hardest user-facing work, fine-tuned open weights for everything high-volume, private, cost-sensitive or offline — chosen workload by workload.
The key models
The ~22 open models that define the mid-2026 landscape: who builds them, what they activate, what they ship under, and what each one signals. Figures as of 8 July 2026; prices, rankings and specifications are point-in-time snapshots of a fast-moving field.
Profiled models by builder region — July 2026. Click a bar to filter the entries below.
The architectural extremes tell the story. DeepSeek-V4-Pro spans 1.6 trillion total parameters yet activates 49B per token; gpt-oss-120b activates just 5.1B and fits a single 80GB GPU; Llama 4 Scout claims a 10-million-token context window; Kimi K2 Thinking ships a trillion parameters quantisation-aware-trained to INT4. Sparse MoE is now the default above ~100B — and it is why the biggest open models run outside the biggest datacentres.
About Mirai Collective
Mirai Collective is an AI development studio. We build AI products end-to-end, train whole teams from zero to one, and guide organisations through AI adoption — practitioners, not theorists. The full research corpus behind this briefing — the model comparison and hardware/cloud cost sheets, the local-setup quickstart and the source log — is available on request: hello@miraicollective.io