Open-source AI in mid-2026: state of play
Where the open-weights frontier stands halfway through 2026, based on the 55 models we index.
Halfway through 2026, open-weights AI is closer to the closed proprietary frontier than at any prior year. Frontier reasoning is now genuinely open (DeepSeek R1 under MIT, Qwen 3 235B under Apache 2.0, Kimi K2 under modified MIT). The 70B tier is a solved problem for production chat, and 24 GB of consumer VRAM is enough for a 32B model that beats mid-2024 flagships.
This is the review our directory would give if you asked it to summarise the year so far: the specific shifts, the licence changes worth watching, and the practical picks per workload as of July 2026.
The frontier moved twice
Two releases redrew the top of the open-weights leaderboard in 2025 and into early 2026. DeepSeek R1 (January 2025) proved that RL-trained reasoning at ~671B MoE could be trained for a fraction of the compute that OpenAI's o1 required. Its MIT licence meant even the weights had no strings — a first for a model this capable. Six months later, Kimi K2 pushed frontier reasoning and tool use further with a 1T-parameter MoE trained on 15.5T tokens.
In parallel, Meta shipped Llama 4 (April 2025), moving the Llama family to mixture-of-experts with two flagship variants: Scout (17B active / 109B total, 10M context) and Maverick (17B active / 400B total, 1M context). Both are natively multimodal — vision is now standard rather than a separate model.
70B is a solved problem
If your workload needs frontier-adjacent quality without hosting hundreds of billions of parameters, the 70B tier is the answer. Llama 3.3 70B remains excellent; Qwen 2.5 72B is close and permissively licensed; Nemotron 70B leads Arena-style human preference tests. None of these were possible on a single H100 two years ago.
- Chat + long context: Llama 3.3 70B or Qwen 2.5 72B.
- Reasoning at 70B: DeepSeek R1 Distill Llama 70B — MIT-licensed teacher, Llama-3-licensed student.
- Agent workflows with reliable tool use: Hermes 3 Llama 3.1 70B.
Local inference finally makes sense
A single RTX 4090 runs 32B models at Q4_K_M with room for real context. Apple Silicon at 64 GB unified memory runs 70B at Q4. Qwen 3 32B is our current default recommendation — Apache 2.0, reasoning-mode toggle, strong across chat and code.
The gap with hosted frontier models is smaller than it's ever been. If your workload can tolerate 30–60 tokens/sec instead of hosted-frontier 100+, the cost-per-token math tilts strongly toward self-hosting once you exceed 1B input tokens per month.
Licence tiers matter more than ever
The Llama family shifted from Llama 3 CLA to Llama 4 CLA. Qwen 3 unified under Apache 2.0 including the 235B flagship — reversing the Qwen 2.5 pattern where only sub-72B was Apache. DeepSeek R1's MIT licence gave it a permissive halo that the base V3 (custom DeepSeek licence) doesn't share.
If commercial deployment is on the table, read every licence yourself before committing. Our /licenses pages summarise the terms, but a summary isn't legal advice.
What's next
- Watch for Llama 5 late 2026 or early 2027 based on Meta's cadence.
- Reasoning-mode toggles will be table stakes. Qwen 3 mainstreamed the pattern; expect DeepSeek and Meta to follow.
- SWE-bench Verified is replacing HumanEval as the practitioner-facing coding benchmark. Watch scores climb from ~30 % (Q1 2026) toward 60–70 % by year-end.
For our specific picks per workload, see /best. For the raw catalogue, /models. For hardware and quantization how-tos, /hardware and /quantization.