Two releases landed in October 2026 that both got called "open model" by people who should know better, and they are not the same kind of thing.
Reflection AI shipped Beam on 5 October. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, pretrained on 23.8 trillion tokens in under four weeks across 6,144 NVIDIA GB300 GPUs. Reinforcement learning ran on 10,500 of the same GPUs and generated more than 100 million rollouts. Weights are Apache 2.0, with tools for running and fine-tuning on Hugging Face. The company claims Beam matches GLM 5.2 on advanced reasoning benchmarks at three to four times less inference compute.
On 19 October, at Bali Tech Week, GoTo released GoTo-AI Gemma 3 27B, an instruction-tuned model for English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese with a 128K context window. GoTo's own leaderboard puts it at 64.39 overall against 59.69 for base Gemma 3 27B, with the largest gains, roughly 7 to 10 points, in Balinese and Batak Toba. The Indonesian number moved only slightly.
Both are downloadable. Only one is open source.
What 23 billion active parameters buys you
The interesting number in Beam is not 501B. It is 23B.
A sparse MoE with 23B active parameters needs roughly the memory bandwidth of a 23B dense model per generated token while carrying the knowledge capacity of something much larger. That gap between total size and per-token cost is where Reflection's inference compute claim comes from, and it is why the model is servable by organizations that could never host a 500B dense checkpoint. Serving 23B active is a multi-GPU job, but it is a job a national lab can actually staff.
Whether "matches GLM 5.2" holds outside Reflection's own evals is a separate question. The company has not published independent benchmark runs, and 23.8 trillion tokens in four weeks implies an aggressively tuned training recipe. I would want third-party coding and reasoning numbers before moving a workload onto it. But the shape of the release, Apache 2.0 plus tooling plus a named customer in the US Department of Energy, is a deliberate pitch to buyers who cannot route their data through a US-hosted API and will not route it through a Chinese one either.
Reflection is selling a model you can run in your own rack without asking permission. The model itself is almost secondary to that.
GoTo's 27B model is a different kind of release
A 27B dense model at full 16-bit precision wants 54 to 60GB of GPU memory. The model card suggests one 80GB A100, two L40S cards, or three RTX 4090s. That is beyond a laptop, and for most Indonesian universities and regional government agencies it means cloud GPUs or a quantized build. GoTo has not published quantized files, though community tooling for Gemma 3 will produce them.
The evaluation is thin by frontier lab standards: samples of 100 to 1,000 items per dataset, run by GoTo itself, compared against Qwen3 and Gemma 3 with no newer open models in the mix. Nobody independent has tested it yet.
None of that makes the release useless. A model that is measurably better at Balinese and Batak Toba has value precisely because the large labs have no commercial reason to serve those languages. The gains there are the point. The flat Indonesian score is the honest part of the leaderboard, and I would rather see it published than smoothed over.
Open weights is not open source
Here is where I get annoyed.
GoTo's weights are public and free for commercial use, but they ship under Google's Gemma Terms of Use. Those terms carry a prohibited-use policy that travels to every derivative, and Google reserves the right to restrict use that violates it. That is not an open-source license in any sense the OSI would recognize. Much of the local press described the model as open source anyway.
For a startup, the strings probably do not matter. For a government body with a legal review, or anyone planning to redistribute a fine-tune, they do. Apache 2.0 on Beam is a different legal posture, and Reflection is counting on procurement teams noticing. If you are building a sovereign deployment, the license is part of the technical decision, not paperwork you hand to legal after the model is chosen.
There is also a small trap in the GoTo release worth flagging: the example code on the model card loads a repository named GoToCompany/GoTo-AI-Gemma-3-27B-IT with a comment telling users to update the path. The weights actually live at GoToCompany/gemma3-27b-cpt-gotoai-v3-instruct. Copy the sample and you get a 404 before you get a token.
Practical notes for evaluating either model
Check the license before the benchmark. Apache 2.0 and Gemma Terms are not interchangeable, and the difference surfaces during legal review, not during prototyping.
Assume self-reported numbers. Both releases rest on vendor evals. Beam's parity claim with GLM 5.2 needs an independent run before it goes near production, and GoTo's leaderboard has no independent replication at all.
Budget for quantization. A 27B model at 54 to 60GB in 16-bit will not fit on the hardware most interested parties own. Plan for 4-bit or 8-bit copies and measure the accuracy drop on your own data instead of trusting a model card.
Rerun safety evals in the target language. Recent research found models refuse more harmful requests when told they are being evaluated in Indic languages. Safety behavior in English does not transfer automatically to Javanese or Batak Toba, and nobody has published results for this model in those languages.
The US-China framing around open-weight models is real, and the Elec's reporting on that rivalry is a fair read of the procurement dynamics. But the buyer's decision is simpler than the geopolitics. You either need weights you can inspect and run on your own metal, or you do not. Beam and GoTo's Gemma fine-tune are two different answers to that question, published nine days apart, under two licenses that mean very different things.