The first days of October
Argon, 125B local, Ollaya
Google unveiled Gemini 4 Argon, while Strata brought a 125B MoE to ordinary PCs and Ollaya brought decision models local. We separate vendor numbers from measured results — and note where each one loses.Oct 2, 2026
The first days of October delivered big moves at both ends of the stack. Google announced Gemini 4 Argon, the opening model of its next Gemini 4 series — and in the same window came Strata, an inference engine that runs a 125-billion-parameter model on an ordinary PC, and Ollaya, which runs decision models locally. Here is what is verified, what is vendor-claimed, and what to watch.
- Verified Confirmed in official blogs, GitHub or primary documents
- Vendor claim Announced by the developer or vendor; not yet independently verified
- Independent check Confirmed by outside developers, media or hands-on testers
- AIfrontier view Our interpretation
1. Gemini 4 Argon: Google's frontier retryVendor claim
On September 30, 2026, Google announced Gemini 4 Argon, the opening model of its next Gemini 4 series. Its headline change is an output-token limit raised from 64K to 1 million, aimed at finishing long, complex tasks in a single pass. Source: Google
CEO Sundar Pichai said he "wanted to give an early look as soon as possible" given the discussion around the next model. Thousands of Googlers are reportedly already using it for coding, research and writing. Source: Google AI (X)
| Benchmark | Argon | Comparison | Result |
|---|---|---|---|
| DeepSWE v1.1 (real codebases) | 77.9% | Astra 74.1% / Opus 5.5 74.2% / Fable 5.1 67.4% | Leads |
| Vals Index (enterprise work) | 68.9% | Astra 63.1% / Opus 5.5 67.0% | Leads |
| Vals Finance Agent v2 | 65.4% | Fable 5.1 58.9% / Astra 53.5% | Leads |
| AutomationBench | 51.3% | Astra 41.4% / Opus 5.5 42.5% | Leads |
| CWE-bench v1 (vuln fixes) | 68% | Astra 68% (tie) / Opus 5.5 67% | Tied first |
| FrontierSWE v2 (separate SWE suite) | 55.0% | Astra 65.5% / Opus 5.5 62.3% | Behind |
AIfrontier view Argon leads DeepSWE but trails two rivals by roughly 7–10 points on FrontierSWE v2. Because different suites are built by different groups, "Argon is best everywhere" is premature; independent replication is the missing piece.
Google also detailed internal use. Argon agents analyzed fleet-wide telemetry and autonomously applied memory optimizations, freeing over 300 TiB (with 500 TiB–1 PiB expected in total). For C/C++-to-Rust migrations spanning re2 and libgav1 up to the 800K+-line Fuchsia Zircon kernel, Argon replaced about 32K lines of SIMD code in libgav1 with safe Rust that the compiler auto-vectorizes, producing identical output at 2.7× the speed of the earlier Rust port. Source: 9to5Google Source: Ars Technica
Availability is staged: today it is limited to trusted cyber defenders in the Fairwind Program and internal teams, with a rollout soon to Google AI Ultra subscribers and paid API customers. Introductory pricing is $2 per million input and $10 per million output tokens (cached input 95% off); after the introductory period, $4/$20. Because of its offensive cyber capability, it is being offered to trusted defenders without the usual cyber guardrails.
2. Strata: 125 billion parameters on a PCVendor claim
The URL provided, https://www.strataai.solutions/, is a different company — "Strata AI Solutions," an AI training and consulting firm — and is unrelated to the inference engine discussed here. The engine is an open-source project distributed at GitHub: Niko1221/Strata. A same-named research project on long-context caching (arXiv:2508.18572) is also separate.
Strata is an inference server built specifically to run Qwen3.8-Flash-Next — a 125-billion-parameter mixture-of-experts model that activates only about 6 billion parameters per token — on an ordinary desktop. Requirements: one NVIDIA GPU (12–24 GB VRAM), 64 GB of system RAM and an x86 CPU. It exposes OpenAI- and Anthropic-compatible APIs on localhost. Source: GitHub (release notes)
| Tier | Role |
|---|---|
| GPU (VRAM) | Attention and DeltaNet mixer layers, routers, shared experts, output head, multi-token-prediction layer, KV cache, adaptive cache of hot experts |
| System RAM | All 24,576 experts pinned in memory; AVX-512/AVX2 CPU kernels evaluate experts missing from the GPU cache |
| SSD | A 28.8 GB n-gram table stays on disk, read a few rows per token through the OS page cache |
Published benchmarks claim up to 94.6 tokens/second on an RTX 5070 at 4K context and 65.1 at 128K with the smallest quantization. An early user report cites about 70 tok/s at 128K on an RTX 3090; the project labels its 3090 and 40-series figures as estimates. Source: AlphaSignal
Hands-on checkIndependent check
A developer who installed it reports that on an RTX 5090 (32 GB) with a Ryzen AI MAX+ 395 and ~128 GB RAM on Linux, setup took about 2 hours 15 minutes including an 84 GB model download. At IQ3_S quantization and 128K context, 400-token code generation ran at ~114 tokens/second, and the model solved 96.7% of the coding tasks tried. Source: note
AIfrontier view Three limits are easy to miss: (1) it is single-purpose — built for this one model, not others; (2) independent replication is still thin, and some speed figures are estimates; and (3) the approach leans on this model's MoE sparsity, so it does not automatically transfer to dense giants. The honest reading is "this MoE, on this class of machine," not "any giant model on any PC."
3. Ollaya: decision models, run locallyIndependent check
Ollaya is an app that pulls and runs open decision models locally, in the spirit of Ollama (Apache License 2.0; Windows, macOS, Linux, Docker). It serves TypeSafe-compatible endpoints (/v1/systemone, /v1/decisions, /v1/models), and the official TypeSafe Python SDK 0.7.1 is said to work against it unchanged. Source: Ollaya Source: GitHub
GIGAZINE's example runs the model Winnow-E4B on "This is the third double charge this year. Refund me today or I switch services," returning intent: refund (91%), urgency: yes (92%), anger: 2.89 (86%).
| Item | Ollaya | Ollama 0.35 |
|---|---|---|
| Models | 15 families (encoders: laya, nli, gliclass, von; decoders: winnow, kev, decider, nimble, jeb, jeeves, cygnet, and more) | Nimble and Tev1 (decoders only) |
| Probabilities | Calibrated with each author's temperature; refittable in a Modelfile | Raw softmax (documented as uncalibrated) |
| Options per question | Up to 255 | Up to 26 |
| Questions per request | Up to 256 | Up to 64 |
On the same RTX 4090, winnow:e4b answers a five-question request in 89 ms and scores 0.722 on typed decisions (TypeSafe's hosted Jev: 0.738 at 236–276 ms); laya answers in about 10 ms and runs well on CPU. Ollaya itself concedes that on the same Nimble weights, Ollama is faster. Source: GIGAZINE
4. Jeff, updated: 22 ms and 28 msIndependent check
A September 29 GIGAZINE piece adds detail on Jeff: at 0.8B parameters it makes a decision in about 22 ms on an RTX PRO 6000 and 28 ms on an M4 Max. Three variants exist (Qwen3.5 0.8B, Qwen3.5 2B, Gemma 4 E2B); code is MIT and weights are Apache 2.0. Training ran on a single RTX PRO 6000 (about 2 hours for 0.8B, 3.5 for 2B), with data written by Qwen3.8-Flash-Next on two DGX Sparks and no cloud GPU. It states plainly that it is unaffiliated with, and unendorsed by, TypeSafe AI. Source: GIGAZINE
5. A practical pattern: turn decisions into questionsAIfrontier view
A practitioner's guide making the rounds splits an agent's steps into four buckets. Source: X (polydao)
| Bucket | Owner | Example |
|---|---|---|
| Generation | LLM | Writing, code generation |
| Exact rules | Code | Fixed conditions, validation |
| Choice / judgment | Decision model (e.g. Jev) | Which tool next, is this source relevant, is this command safe |
| Irreversible action | Human | Deploy, send, pay |
The move is to open an agent transcript, count the moments it was not creating anything, rewrite them as questions, and branch on calibrated confidence — escalating only the hard, low-confidence cases to a frontier model (the guide uses Kimi K3). That beats running everything through the largest model on both cost and latency.
6. What it adds up toAIfrontier view
The first week of October moved both ends of the stack at once. Gemini 4 Argon pushes a new ceiling with a 1M-token output limit, but loses some suites to rivals, so its full picture awaits outside testing. Strata and Ollaya, meanwhile, pull server-GPU-scale work down to the desk. They are not competitors but different tiers. Hard, large jobs to frontier APIs; high-volume, routine decisions local — that two-tier setup is becoming the realistic architecture for late 2026. If you are evaluating any of it, measure on your own data and your own machine: vendor numbers only take you so far.
Note: the URL supplied for Strata (strataai.solutions) belongs to an unrelated company; this article is based on the project's GitHub repository and third-party coverage. Many benchmark figures are vendor-published, with limited independent replication.
Sources
- Google「Gemini 4 Argon」公式ブログ https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- Google AI（X）の発表 https://x.com/GoogleAI/status/2105388478683119904
- 9to5Google「Gemini 4 Argon announcement」（9/30） https://9to5google.com/2026/09/30/gemini-4-argon-announcement/
- Ars Technica「Google announces Gemini 4 Argon…」（9/30） https://arstechnica.com/google/2026/09/google-announces-gemini-4-argon-ai-model-but-you-cant-use-it-yet/
- GIGAZINE「GoogleがGemini 4 Argonを発表」（10/1） https://gigazine.net/news/20261001-google-gemini-4-argon/
- GitHub「Niko1221/Strata」リリースノート https://github.com/Niko1221/Strata/releases
- AlphaSignal「Strata serves a 125B MoE from a consumer PC」 https://alphasignal.ai/news/strata-runs-a-125b-ai-model-on-a-regular-gaming-pc
- note「Strataを2日間導入してみた」 https://note.com/zephel01/n/nce0d0c3328cd
- Ollaya公式サイト／GitHub（ollaya-dev/ollaya） https://ollaya.dev/
- GIGAZINE「Ollaya」（10/1） https://gigazine.net/news/20261001-ollaya-run-decision-model/
- GIGAZINE「Jeff」（9/29） https://gigazine.net/news/20260929-jeff/
- X（polydao）エージェント設計の実務ガイド https://x.com/polydao/status/2104783226833186920