Open Source AI Models That Rival GPT and Claude: DeepSeek, Qwen, Kimi K2, and GLM

Alibaba’s Qwen family now accounts for more than half of all open-weight model downloads worldwide, according to Hugging Face data reported by the South China Morning Post. Not Meta’s Llama. Meta is the company that made “open-weight frontier model” a category people cared about, and as of 2026 it isn’t in the conversation anymore.

That’s the part of this story that surprised me most, so it’s worth starting there before getting into who’s actually winning.

Where Meta went wrong

Llama 4 landed in April 2025 to accusations that Meta had tuned a special benchmark version rather than the one it shipped — LMArena partly corroborated this — and independent replications couldn’t reproduce Meta’s own claims that it beat GPT-4o, Gemini 2.0, and DeepSeek V3.1. On one widely cited coding benchmark, Llama 4 finished last among the models tested, 18 points behind GPT-4o. By mid-2026, 11 of the 14 researchers on Meta’s original 2023 Llama paper had left the company. Meta hasn’t shipped a credible open-weight update since.

Into that vacuum stepped four labs, three of them Chinese, plus — unexpectedly — OpenAI itself.

The current open-weight frontier

DeepSeek V4-Pro went GA on August 13, 2026: 1.6 trillion total parameters, 49 billion active per token (mixture-of-experts), MIT-licensed weights on Hugging Face the same day it launched. It posts 80.6% on SWE-bench Verified — the strongest published open-weight result on that benchmark — and 90.1% on GPQA Diamond.

Qwen3.8-Max, Alibaba’s flagship, scales to 2.4 trillion parameters and is multimodal, released under Apache 2.0. Earlier in the Qwen3 line, the 235B-A22B model already beat Claude Opus 4 and GPT-4o on 14 of 25 published benchmarks — Qwen’s edge has been consistency across a wide benchmark spread rather than any single headline number.

Kimi K2, from Moonshot AI, is a 1-trillion-parameter MoE model (32B active) released under a modified MIT license — the only catch is you have to display “Kimi K2” in your UI if your product clears 100 million monthly users or $20M in monthly revenue. Moonshot has iterated fast: K2 Thinking added reasoning in November 2025, K2.5 added native vision in January 2026, K2.6 shipped in April 2026. Its specialty is agentic workloads — it leads SWE-Bench Pro among open models, the benchmark built specifically around multi-step tool use rather than single-shot coding.

GLM-5, from Zhipu AI, released February 11, 2026 under MIT: a 744B-parameter MoE with roughly 40B active. Its later point release, GLM-5.2, scores 62.1 on SWE-bench Pro against GPT-5.5’s 58.6 on the same benchmark — an open-weight model outright beating a frontier closed model on a real coding benchmark, not just approaching it.

Then there’s gpt-oss, OpenAI’s own August 2025 release — its first open-weight models since GPT-2 in 2019, also Apache 2.0. gpt-oss-120b (117B total, 5.1B active) sits between o3-mini and o4-mini on OpenAI’s own internal comparisons, and the smaller gpt-oss-20b holds up surprisingly well against it despite being six times smaller. The company that arguably has the most to lose from open weights being good enough shipped some anyway.

How close is “comparable,” actually

Artificial Analysis’s Intelligence Index, which aggregates a wide benchmark suite into one number, put GPT-6 Astra and Claude Fable 5.1 at 53 as of September 2026. The best open-weight models — GLM-5.3 and Kimi K3, the next releases after the ones above — sit at 44. A nine-point gap, on a scale where the field’s whole range runs roughly zero to sixty.

That gap isn’t even across task types. It’s narrowest on structured tasks — code, math, formal reasoning — where GLM-5.2 beating GPT-5.5 on SWE-bench Pro isn’t a fluke, it’s representative. It’s widest on creative writing and the kind of nuanced instruction-following that’s hard to reduce to a benchmark score at all. If your workload is “write code, prove a theorem, run an agent loop,” open weights are at parity or better today. If it’s “write something a person enjoys reading,” the frontier closed models still have real room on open models.

Why the gap closed this fast

Three things, and only one of them is technical. First, MoE architectures got good enough that a trillion-parameter model can run with 30-50B active parameters, which makes serving these things at all economically feasible — DeepSeek, Kimi, and GLM are all MoE for exactly this reason. Second, three well-funded Chinese labs decided open weights was the right distribution strategy for gaining users and mindshare against Western closed labs that won’t sell them frontier API access as cheaply, and they iterate on a roughly monthly cadence now, not annually. Third, and this is the one Meta’s absence makes obvious: releasing open weights is not something you get right by accident. DeepSeek, Alibaba, Moonshot, and Zhipu are matching or beating frontier closed labs specifically because they’re treating it as their main strategy rather than a side project or a PR move.

One terminology note, since it matters if you’re picking a model for something that touches licensing: none of these are “open source” in the sense of the training data or training code being public. They’re open-weight — the trained parameters are downloadable and, under MIT or Apache 2.0, usable and modifiable freely. That’s the term researchers actually use, “open source” is just what most people search for and what stuck.

Sources