Open-weight or closed AI models: what companies should weigh in 2026
· TexaHub · Artificial Intelligence, LLMs & Foundation Models, Model Comparisons, Technology Trends
Open-weight models now trail closed frontier systems by months rather than years, but licences, hardware and safety work decide whether they suit a given deployment.
Where the rankings stand
On Artificial Analysis's open-weights comparison page, accessed on 1 October 2026, the highest-scoring open models on its Intelligence Index are Xiaomi's MiMo-V2.6-Pro at 46, Z.ai's GLM-5.3 (Max) at 45, Moonshot's Kimi K3 (Max) at 44, GLM 5.3 Flash at 42 and DeepSeek V4.1 Flash (Max) at 39 [1]. The same page says the index, version 4.3.2, combines 10 evaluations, including Terminal-Bench 4.0, SciCode and Humanity's Last Exam [1]. On the evaluator's main leaderboard, viewed the same day, Claude Opus 5.5 (max with fallback) leads with 58, while GPT-6 Astra (max) and Gemini 4 Argon (high) each score 53 [2].
The crowd-voted Arena text leaderboard tells a similar story from a different angle. In its 30 September 2026 update, built on 8,602,501 votes, Gemini 4 Argon (high) ranks first with a score of 1525 [3]. The highest-placed open-weight model is kimi-k3-max at rank 17 (1488), followed by mimo-v2.6-pro at rank 25 (1480) and glm-5.3-max at rank 26 (1479) [3].
Snapshots move quickly. Artificial Analysis's 30 April 2026 review had Kimi K2.6 and MiMo V2.5 Pro tied as the leading open models at 54, with GPT-5.5 (xhigh) leading proprietary models at 60 [4]. The scores in that review are much higher than those on the current page, so figures from different snapshots should not be compared directly. Read the date of every ranking alongside the score.
How large is the gap?
Estimates of the lag vary with who is measuring. The US Center for AI Standards and Innovation (CAISI) evaluated DeepSeek V4 Pro in May 2026 and concluded that its capabilities 'lag behind the frontier by about 8 months' [5]. CAISI also found that DeepSeek V4 'performs similarly to GPT-5' on evaluations that include non-public benchmarks, even though DeepSeek's own published results suggested parity with much newer models [5]. Writing on 21 September 2026, Nathan Lambert of Interconnects estimated that Chinese open-weight models are 'approximately 2-5 months behind the closed American frontier', and that American open-weight models are about 6-9 months behind OpenAI and Anthropic [6].
On our reading, the gap between vendor-reported scores and independent results is the more useful lesson for buyers. Model cards routinely compare a new release with named closed models, and those comparisons are not independent evaluations. Kimi K3's card, for example, reports 93.5 on GPQA Diamond against 91.0 for Claude Opus 4.8 [8]. Qwen3.8-27B's card reports 61.7 on SWE-bench Pro against 53.4 for Opus 4.6 Max [11].
Price is where open models clearly compete. In April, Artificial Analysis said nine of the 13 models on its intelligence-versus-cost Pareto frontier were open-weight, and that the leading open models offered 'comparable intelligence to leading proprietary models at between half to one-sixth of the price' [4]. CAISI found DeepSeek V4 more cost-efficient than GPT-5.4 mini on 5 of 7 benchmarks [5]. VentureBeat reports that Xiaomi's API charges $0.435 per million uncached input tokens and $0.87 per million output tokens for MiMo-V2.6-Pro, compared with $5.00 and $25.00 for Claude Opus 5 [7].
The families worth knowing
Chinese labs lead the open tier, and Lambert reports that Chinese models now account for over 80% of open-model usage on OpenRouter [6]. Kimi K3 is a mixture-of-experts model with 2.8T total and 104B activated parameters and a context window of 1,048,576 tokens [8]. DeepSeek V4 Pro has 1.6T parameters, 49B of them activated, and supports a context of one million tokens [12]. VentureBeat reports that MiMo-V2.6-Pro has 1.02 trillion total parameters and 42 billion active [7]. Alibaba's Qwen3.8-27B is a much smaller 27B model with a native context of 262,144 tokens [11].
Western labs mostly offer smaller open models. Google released Gemma 4 on 2 April 2026 in four sizes, up to a 26B mixture-of-experts model and a 31B dense model [13]. Mistral Large 3 has 675B total and 41B active parameters, alongside Ministral 3 models at 3B, 8B and 14B [14]. OpenAI's gpt-oss family, released in August 2025, consists of a 120B model (5.1B active) and a 20B model (3.6B active) [15]. Meta, whose Llama models once led the open tier, released Muse Glimmer, an open model of about 29.6 billion parameters, on 10 August 2026 [17]. According to VentureBeat, Mark Zuckerberg has said Meta will also release the weights of Muse Spark 1.2 [17]. On Artificial Analysis's index, Qwen3.8 27B (Xhigh) scores 34, Muse Glimmer (High) 17 and Mistral Medium 3.5 14 [1].
Licences are not all alike
Commercially, 'open-weight' covers very different terms. DeepSeek V4 Pro is released under the MIT License [12], and VentureBeat reports that MiMo-V2.6-Pro is MIT-licensed as well [7]. Qwen3.8-27B uses Apache 2.0 [11], as do Gemma 4 [13], the Mistral 3 family [14], gpt-oss [15] and Muse Glimmer [17].
Other leading models come with custom licences. The Kimi K3 License requires a separate commercial agreement with Moonshot AI before a company operating a Model-as-a-Service business can use the model, if that company and its affiliates have revenue of more than 20 million US dollars over any consecutive 12 months [9]. Products with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, must display 'Kimi K3' prominently in their interface [9]. The GLM-5.3 licence says that Model-as-a-Service operators with revenue above 10 billion US dollars over any 12 consecutive months 'must pass Z.AI's security review' before commercial use [10]. The sources disagree here: Arena lists glm-5.3-max as MIT [3], but the licence file in Z.ai's repository is a custom GLM-5.3 licence [10].
Llama 4's community licence requires companies with more than 700 million monthly active users to request a licence from Meta [16]. It also requires licensees to display 'Built with Llama' and to follow an acceptable use policy [16]. Legal teams should read the licence file itself, not the label shown on a model page.
Hosting, control and fine-tuning
Hardware needs vary widely. The gpt-oss-120b model fits on a single 80GB GPU, and the 20B model runs on 16GB GPUs [15]. Meta says Muse Glimmer needs more than 55GB of memory at full precision, and that roughly 4-bit versions bring its language-model weights under 20GB [17]. Kimi K3 ships with MXFP4 weights and MXFP8 activations and is meant to run on inference engines such as vLLM and SGLang [8]. Mistral lists deployment options ranging from Blackwell NVL72 systems to A100 and H100 nodes and edge devices [14].
In our assessment, the trillion-parameter flagships need data-centre hardware. For many companies, the practical choice is between renting that capacity and running a smaller open model on hardware they already own, rather than between open and closed models in the abstract. Self-hosting keeps prompts and outputs on infrastructure the company controls, and open weights let teams fine-tune the model directly. Closed APIs instead tie fine-tuning and data handling to the vendor's terms. Support is one area where open weights do not remove the need for a vendor: Mistral, for example, sells custom training services to adapt its models [14]. We found little independent evidence that compares the total cost of ownership of self-hosting with paying for APIs, and companies should model their own workloads rather than rely on per-token comparisons.
Safety is the buyer's job
Once weights are published, any safeguards travel with the model and cannot be updated centrally. In a joint assessment published on 23 July 2026, the UK AI Security Institute and CAISI found that 'Kimi K3's safeguards did not prevent it from attempting cyber exploit development or offensive cyber operations' [18]. The same assessment found Kimi K3 well behind frontier models on offensive tasks. It achieved arbitrary code execution on none of 41 ExploitBench samples, against an average of 20 of 41 for the most capable models. In a simulated corporate network attack, it reached step 17 of 32, compared with 28.5 for the leading US models [18].
There is no single winner. Closed models still lead the independent rankings cited here, while open-weight models offer lower prices, local control and, in several cases, permissive licences. Companies adopting open weights take on the work of testing, red-teaming and monitoring that an API provider would otherwise share. Companies choosing closed models accept dependence on the vendor in exchange for higher top-end scores.
Sources
[1] Comparison of Open Source AI Models across Intelligence, Performance, Price, Context Window, and more — Artificial Analysis, accessed 1 Oct 2026
[2] LLM Leaderboard - Comparison of AI models from OpenAI, Anthropic, Google, SpaceXAI & others — Artificial Analysis, accessed 1 Oct 2026
[3] LLM Leaderboard - Best Text & Chat AI Models Compared — Arena (formerly LMArena), updated 30 Sep 2026
[4] Recent Open Weights Model Launches — Artificial Analysis, 30 Apr 2026
[5] CAISI Evaluation of DeepSeek V4 Pro — NIST Center for AI Standards and Innovation, 1 May 2026
[6] The current balance of power in open models — Interconnects (Nathan Lambert), 21 Sep 2026
[7] 'Better than DeepSeek': Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model in the world alongside cheaper V2.6-Flash — VentureBeat (Carl Franzen), 21 Sep 2026
[8] moonshotai/Kimi-K3 — Moonshot AI on Hugging Face, accessed 1 Oct 2026
[9] LICENSE · moonshotai/Kimi-K3 — Moonshot AI on Hugging Face, accessed 1 Oct 2026
[10] LICENSE · zai-org/GLM-5.3 — Z.ai on Hugging Face, accessed 1 Oct 2026
[11] Qwen/Qwen3.8-27B — Qwen (Alibaba) on Hugging Face, accessed 1 Oct 2026
[12] deepseek-ai/DeepSeek-V4-Pro — DeepSeek on Hugging Face, accessed 1 Oct 2026
[13] Gemma 4: Byte for byte, the most capable open models — Google, 2 Apr 2026
[14] Introducing Mistral 3 — Mistral AI, 2 Dec 2025
[15] Welcome GPT OSS, the new open-source model family from OpenAI! — Hugging Face, 5 Aug 2025
[16] Llama 4 Community License Agreement — Meta, licence version 5 Apr 2025
[17] Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter AI model optimized for agents — VentureBeat (Carl Franzen), 10 Aug 2026
[18] UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities — NIST Center for AI Standards and Innovation, 23 Jul 2026 (updated 28 Aug 2026)