Five frontier-class open models dropped in a single month. Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1. If you run a service business and you’re paying top dollar for API calls to closed models, this wave of open model releases just changed your math.
The speed here is the story. A year ago, open models were clearly a tier below the best closed options from OpenAI and Anthropic. That gap is shrinking fast, and some of these new releases come with Apache 2.0 licenses, meaning you can deploy them locally, modify them, and use them commercially with zero licensing headaches.
What happened
- Google released Gemma 4 with dense models at 4B, 9B, and 31B parameters, plus a 26B mixture-of-experts (MoE) variant. The big deal: Google switched to an Apache 2.0 license, removing all the legal gray areas that came with previous custom licenses.
- DeepSeek dropped V4 in two sizes. Pro is a massive 1.6 trillion parameter MoE model. Flash is a leaner 284B model. Early reports say Flash is actually the star, punching above its weight class while Pro underdelivers relative to its size.
- Moonshot AI shipped Kimi K2.6, an update that delivers stronger performance across the board and focuses on long-horizon tasks (think: AI that runs for hours to complete complex work).
- Xiaomi released MiMo 2.5 Pro under Apache 2.0. This is remarkable. A year ago Xiaomi was just getting started with open models. Now MiMo 2.5 Pro trades blows with Kimi K2.6 and GLM-5.1 in both benchmarks and real-world usage.
- GLM-5.1 arrived from Zhipu AI with improved scores across the board and a focus on long-horizon task completion.
- Bonus releases: Poolside AI’s Laguna XS.2 (a 33B coding-focused model), Qwen 3.6, and Arcee AI’s Trinity-Large-Thinking also launched this month.
The numbers
- The open-to-closed model gap has stayed roughly 3 to 7 months since DeepSeek R1, according to Epoch AI’s ECI measurement.
- CAISI (Center for AI Standards and Innovation) ran evaluations across 9 different benchmarks and concluded open models lag behind the American frontier, with the gap widening. But there’s a catch (more on that below).
- DeepSeek V4 Flash uses a 284B parameter, 13B active MoE architecture, meaning it runs much cheaper than its total parameter count suggests.
- LiquidAI’s LFM2.5 was trained on 28 trillion tokens for just 350M parameters, possibly the most overtrained model ever released.
- Poolside’s Laguna XS.2 runs at 33B total, 3B active parameters, making it practical for local deployment on consumer hardware.
5 open models worth testing this week
- Gemma 4 26B-A4B (Google). Best all-around pick for service businesses. Apache 2.0 license means no legal surprises. The MoE architecture keeps inference costs low. If you need a general-purpose model you can actually deploy on your own servers, start here.
- DeepSeek V4 Flash. The performance-per-dollar champion. At 13B active parameters inside a 284B total, it delivers strong results while staying cheap to run. The detailed tech report covers architectural changes for better long-context performance, which matters for document-heavy workflows.
- Kimi K2.6 (Moonshot AI). The long-horizon specialist. If you need AI that can grind through multi-hour tasks like research, code migration, or document review, this one’s built for it. Open models that can sustain quality over extended runs are still rare.
- MiMo 2.5 Pro (Xiaomi). The underdog pick. Apache 2.0 licensed, competitive with models from bigger labs, and backed by a company with serious hardware resources. Worth benchmarking against your current stack.
- Laguna XS.2 (Poolside AI). The coding specialist. If your business builds or maintains software, this 33B model runs locally and focuses specifically on code generation. Poolside’s blog post about reward hacking during coding evaluations is worth reading too.
The hot take
The CAISI report saying open models are falling further behind closed ones is misleading. Their benchmarks use basic evaluation setups (a bash terminal and a fixed token budget) instead of the actual tools these models are trained to work with, like Claude Code or OpenCode. That’s like testing a race car on a dirt road and concluding it’s slow. When you evaluate open models with the right harnesses and model-specific prompting, the gap narrows significantly. For most real business tasks (not PhD-level science problems), the best open models are now good enough. And “good enough” at a fraction of the cost, running on your own hardware, with full data privacy? That’s not second place. That’s a better deal.
The Agency OS play
This week, pick one workflow in your business that currently runs on a closed API (GPT-4, Claude, Gemini Pro) and benchmark it against Gemma 4 or DeepSeek V4 Flash. Don’t guess. Actually run 50 real inputs through both and compare output quality. You might find the open model gets you 90% of the quality at 20% of the cost. For document-heavy work like contract review, intake processing, or report generation, that tradeoff is almost always worth it.
If you handle sensitive client data (and if you’re in legal, healthcare, finance, or HR, you do), the Apache 2.0 models deserve special attention. Gemma 4 and MiMo 2.5 Pro both carry that license. That means you can run them on-premises or in a private cloud without sending a single byte to a third-party API. Set up a local inference server this month. Tools like Ollama, vLLM, or llama.cpp make this surprisingly straightforward, even on modest hardware, for the smaller model sizes.
Finally, don’t sleep on the long-horizon capabilities of Kimi K2.6 and GLM-5.1. If you’ve been hesitant to let AI agents run autonomously on multi-step tasks because the quality degraded after a few minutes, test again. These models are specifically optimized for sustained performance over hours. That opens up workflows like automated research reports, competitive analysis, and large-scale data processing that simply weren’t reliable with earlier open models.