September 2026 has settled into a pattern: the most interesting AI launches are no longer about bigger models in bigger clouds, but about capability moving out of the cloud and into your hands: onto your own GPU node, your laptop, your phone, your robot, your desktop. The best new AI tools September 2026 produced follow that pattern: an open-weight security model you can air-gap (Aikido Altar-1), a 2.6-billion-parameter agent that runs on a phone (Liquid AI’s LFM2.5-2.6B), a $29,990 humanoid robot with an Apache-2.0 SDK (Feather Robotics), an MIT-licensed canvas that makes coding agents legible (Whiteboard 0.1), and a GNOME proposal to fix developer tooling on immutable Linux (Toolpak).
One note before we dig in: Toolpak is a proposal with a funded prototype, and Whiteboard is at 0.x — that’s exactly why they’re worth watching. For each of the best new AI tools September 2026 rounded up below: what it does, why it matters, quick-start notes, and license/pricing — all from primary sources fetched for this post.
Aikido Altar-1 — an open-weight security model you can air-gap
What it does
Altar-1 is security company Aikido’s first open-weight model: a heavily compressed build of Z.AI’s GLM-5.3, tuned for defensive security work. GLM-5.3 is Z.AI’s 753-billion-parameter mixture-of-experts model where each token uses 8 of 256 expert sub-networks per layer (roughly 40 billion active parameters). Altar-1 applies two compression steps. First, REAP (Router-weighted Expert Activation Pruning, a Cerebras Research technique) scores each expert’s contribution and deletes the least useful ones — no retraining — keeping 168 of 256 experts per layer. Then the routed experts are quantized to INT4 via AWQ (W4A16), while attention, the shared expert, and dense layers stay BF16. The result: 328 GB of weights, down from 1,506.7 GB at full precision — a 78.2% reduction, per Aikido’s announcement, with a KL divergence of just 0.506 nats versus the full BF16 parent.
Does the compression hurt security capability? Aikido ran Altar through its internal CVE benchmark — 32 known vulnerabilities across 30 repositories, three runs per case. Altar averaged 60.4% recall and rediscovered 23 of the 32 at least once; the AWQ-quantized parent averaged 61.5% with the same 23, the full-precision parent 65.6% with 25. So Altar retains 92% of its parent’s vulnerability coverage at a third less storage than the already-quantized parent.
In practice, Altar powers Aikido Machine, Aikido’s autonomous pentesting appliance, which runs entirely inside a customer’s own infrastructure, including fully air-gapped environments. Per Aikido’s blog, Altar was deployed to the fleet immediately after evaluations and “identified a valid critical-severity vulnerability during a client production pentest” shortly after. The calibration traces (pentesting harness data plus multilingual Wikipedia — no customer data involved, per the blog) and the fidelity study are public. If you’ve followed the push toward running agentic AI locally on your own hardware, Altar-1 is the first open-weight security model that fits that deployment model.
Why it matters now
Closed frontier models run on someone else’s infrastructure. For a bank under a data-residency mandate, a hospital group with strict data-processing rules, or an industrial operator whose OT environment has no route to the internet, sending source code and unremediated findings to a third-party inference service isn’t a trade-off they’re allowed to make. A frontier-grade security model that fits on a single 4×H200 node and can be served behind an air gap is a genuinely new option — and the methodology, calibration traces, and fidelity numbers are all public.
Quick-start / install notes
From the model card, serving on a Hopper-class node (H100/H200) is one command:
1 | vllm serve aikido/altar-1 --tensor-parallel-size 4 --trust-remote-code --max-model-len 131072 |
(Note: the model card’s literal command references aikido/altar-1, while the Hugging Face repo id is AikidoSec/altar-1.)
You’ll need Hopper-class GPUs: 328 GB of weights across 4×H200 leaves room for a 128k-context KV cache at production batch sizes, and vLLM auto-selects the Marlin MoE kernel.
License & pricing
The Altar-1 model card states it inherits the GLM-5.3 license. I fetched that license from the Z.AI GLM-5.3 repository: it grants free use, modification, distribution, sublicensing, and selling of the weights and code. One notable condition: Model-as-a-Service businesses with aggregate revenue (with affiliates) over $10 billion in any consecutive 12 months must pass Z.AI’s security review before commercial use. For everyone else, it’s effectively free weights.
Liquid AI LFM2.5-2.6B — an agentic model small enough for a phone
What it does
One honesty note up front: LFM2.5-2.6B is technically an August 2026 release — Liquid’s own blog dates the announcement to August — but its tool-calling benchmark leads are exactly the kind of on-device milestone this roundup tracks, so it earns a September slot.
LFM2.5-2.6B is a 2.69-billion-parameter dense hybrid model purpose-built for agentic workloads on-device: planning, calling tools, and working through multi-step tasks without any cloud API. It was pre-trained on about 34 trillion tokens, has a 131,072-token context window, and a vocabulary doubled to 128,000 to support 16 languages.
The interesting engineering is in the post-training pipeline described in Liquid AI’s release post: supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, and finally agentic reinforcement learning run inside real agent harnesses — including Hermes Agent and OpenClaw — so the model learns the tools and interaction patterns of popular agent environments. The result is native tool calling, writing Pythonic function calls between special tokens by default.
The benchmark story is strong for its size. Liquid’s blog reports LFM2.5-2.6B leads every instruction-following benchmark in its comparison set (IFBench 59.17, Multi-IF 80.07) and nearly every tool-use benchmark (ToolSandbox 77.83), trailing only Qwen3.5-9B on BFCLv4, while competing against models up to four times larger. Coding is the one area where bigger models keep an edge — our DeepSeek V4 developer guide covers that side of the fence — and the model card is refreshingly blunt about scope: recommended for agentic workflows, tool use, extraction, RAG, and long-context work; not for agentic coding and knowledge-heavy tasks.
The speed numbers are the real story: 220 tokens/s on an Apple M5 Max and 113 tokens/s on a Ryzen AI Max+ 395 while staying under 2.5 GB, and 30 tokens/s on a phone. On GPU it’s the fastest model in its size class in Liquid’s tests — almost 15K output tokens per second at high concurrency on a single H100. Free inference plus that throughput is the pitch: agents running in parallel around the clock at no marginal cost. (For more on small open models for edge inference, see our guide to how to run Gemma 4 locally.)
Why it matters now
Edge agentic inference has been a demo, not a product. What’s changed is that a 2.6B model can now hold a 128K context, call tools reliably, and run inside the same harnesses people use in production — all locally. For privacy-sensitive automation or anyone burning millions of tokens on background agents, the economics flip completely.
Quick-start / install notes
Both base and post-trained models are on Hugging Face (LiquidAI/LFM2.5-2.6B and LFM2.5-2.6B-Base), with day-one checkpoints for llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX. The fastest path from the docs: serve the model behind an OpenAI-compatible endpoint, then point your agent harness at it — the model card includes worked configs for Hermes Agent, OpenClaw, and Pi, plus a browser-based Research Agent demo. Bonus: the companion LFM2.5-2.6B-DSpark drafter (328M) adds roughly 2.6× faster decoding with identical outputs.
License & pricing
The weights are free, under the LFM Open License v1.0 (I pulled the LICENSE file from the model repository). It grants a perpetual, royalty-free license to use, modify, and redistribute the weights and derivative works. There is a Commercial Use Limitation: commercial use is conditioned on your legal entity not exceeding the license’s Threshold, defined in Section 5 as $10 million ($10,000,000) in annual revenue — exceed that and commercial use isn’t licensed, and Section 11 terminates the license on non-compliance. Liquid’s blog calls the family open-weight — “download, fine-tune, and deploy” — but companies above $10M in annual revenue should read Section 5 before shipping.
Feather Robotics — the $29,990 humanoid that wants to be the Android of robotics
What it does
Feather is a bimanual wheeled mobile manipulator built for developers: two 7-degree-of-freedom arms with 1-meter reach, a holonomic omnidirectional base, and 23 DoF total. The hardware page is unusually specific — 105 kg, a 500×510 mm footprint that fits through standard doors, 600 mm of vertical travel, dual hot-swappable 48 V batteries rated for 10 hours under a 240 W load, Zed stereo cameras, a swappable Jetson computer, and a 4 lb rated sustained single-arm payload (15 lb peak hold per arm).
The software story is the platform play: an open Python 3.9+ SDK, licensed Apache 2.0, and bring-your-own-AI — the hardware runs models from any leading robotics AI provider, including Nvidia, Skild, or Physical Intelligence, per TechCrunch’s profile of the company.
Why it matters now
The humanoid market splits into consumer-helper moonshots and frontier general-purpose bets; Feather is deliberately neither. As co-founder Hoa Mai told TechCrunch, “You can’t buy a Tesla robot today and develop on top of it” — so Feather sells a deployable robot at a developer-reachable price and expects the ecosystem to build the applications. There’s traction behind the pitch: a $7.6 million pre-seed led by Gradient Ventures, over $1 million in revenue, and robots already working as restaurant cooks and lab cleaners. And at $29,990 — roughly half the price of comparable hardware — it undercuts the field by a wide margin.
Quick-start / install notes
Start at feather.dev — the technology page doubles as the full spec sheet, and the SDK documentation covers getting your model to a physical deployment. Note the quick-start here is for software: the robot itself is a purchase, not a download.
License & pricing
Straight from feather.dev’s own table: the starting price is $29,990, a limited-time promotional price (down from the $32,990 list), with a 1-year warranty and extended SLAs for fleets of 5 or more. The SDK is Apache 2.0 — open source, per the vendor’s spec table.
Whiteboard 0.1 — a shared canvas that makes coding agents reviewable
What it does
Whiteboard, from the /dev/fast team, is an open-source desktop app “where humans and agents can architect software together in a common workspace,” per the project README. It plugs into the coding agents you already use — Claude Code, Codex, and others — and gives them an SDK to draw on an in-app canvas: sequence diagrams, entity-relationship diagrams, quotes from the agent’s trace (source: devdotfast/whiteboard on GitHub). Click any visualization and you jump to the underlying code, with VSCode keybindings and LSP support. (If agent-written code is your daily reality, see our deep dive on DeepSeek ReasonX, an open-source terminal AI coding agent.)
Three features make it more than a pretty diagram tool. There’s a semantic, AST-aware diff viewer written in Rust that filters raw diffs to relevant changes — large added functions summarized as pseudocode, tests and docs collapsed by default, customizable via a WASM plugin system. There’s a decision log, letting agents query and link their own traces so you can see which requirements they followed and which decisions they made autonomously. And it’s built on a vendored Code-OSS fork, chosen because — as the README wryly notes — “we only use our text editors for reviewing line-by-line diffs now.”
The current limitations are honestly listed: no file editing yet, weak multi-repo support, and shared reviews that don’t propagate updates. The desktop package sits at 0.1.3.
Why it matters now
As more code gets written by agents, the bottleneck shifts from generating changes to understanding them. Whiteboard attacks exactly that: making an agent’s reasoning and output legible enough for a human to review with confidence. It runs against your local checkouts, and its anonymous telemetry explicitly excludes your code, diffs, prompts, and model output.
Quick-start / install notes
Three steps from the README: download the app from install.dev.fast (macOS, Windows, Ubuntu, Fedora), connect Claude Code, Codex, or another coding agent from the welcome screen, then ask your agent to review your current branch against main and open the result in Whiteboard. The team recommends GPT-6 Sol and Claude Opus 5.5 for the intelligence/cost/speed tradeoff. Source is on GitHub at devdotfast/whiteboard.
License & pricing
MIT-licensed (the repo’s LICENSE and package metadata both confirm it), free and open source, with the vendored Code-OSS fork retaining Microsoft’s MIT license. No paid tier exists today.
GNOME Toolpak — a cleaner way to ship dev tools on immutable Linux
What it does
Toolpak isn’t a product yet — it’s a design proposal from GNOME contributor alatiera for how developer tooling should work on image-based Linux desktops. The problem is real: Flatpak solved app distribution on immutable systems, but the developer story never caught up, since apt install doesn’t work when your OS is an image. The post walks through each workaround’s failure mode: rpm-ostree overlays can break the system; GNOME OS’s monolithic “developer overlay” covers only the toolchain for building the OS itself; Toolbox and distrobox replicate package management inside containers tools don’t expect to run in; Homebrew binaries can override and break system binaries — disqualifying it for system development, in the author’s view; and Flatpak’s portal-based sandbox is, by design, wrong for CLI tools like strace or qemu.
The proposed design is opinionated and specific: tools ship as discoverable disk images using the UAPI.3 specification, bringing “state of the art security practices, like Verity” and a path to reproducible builds. Each tool gets its own mount namespace, borrowing Flatpak’s /usr (shared runtime) and /app (tool contents) split. Tools bundle all their dependencies and get unrestricted access to the system, so utilities that can’t realistically run confined don’t need rewrites. A shim prepends tool binaries to your PATH and sets up the namespace on execution. Distribution runs through a reviewed, signed catalog; Flatpak-packagable apps would be rejected.
Why it matters now
Immutable, image-based desktops are winning on security and reliability, but they’ve pushed developers back toward fragile workarounds. If Toolpak’s design holds up, developer utilities could become first-class citizens on immutable systems — installed from a trusted catalog, sandboxed at the image level, yet with the full system access real debugging tools need. (For related Linux developer-workflow coverage, see our Podman vs Docker comparison.)
Quick-start / install notes
There is none, and the author is upfront about it. This is a design-stage proposal with “concrete work towards a prototype as part of a Prototypefund project”; more is promised “in the coming weeks.” Feedback is open in #gnome-os:gnome.org on Matrix or via blog comments.
License & pricing
Not applicable — Toolpak is a proposal and prototype within the open GNOME ecosystem, not a shipped product with a license or price.
Which one for what: a quick guide to the best new AI tools September 2026
So, of the five best new AI tools September 2026 delivered, which one maps to your stack? Here’s the short version:
- Security team with data-residency or air-gap requirements → Aikido Altar-1. If your threat model includes “our findings never leave the building,” this is the only option here built for exactly that — provided you can field 4×H200.
- Edge/agent builder on constrained hardware → LFM2.5-2.6B. Best-in-class instruction following and tool use at 2.6B parameters, 30 tokens/s on a phone, free inference. Skip it for coding-heavy agents — Liquid says so itself.
- Physical-AI developer → Feather. The only developer-priced humanoid here, with an Apache-2.0 SDK and bring-your-own-model. Budget $29,990 and a garage.
- Team reviewing agent-written code at scale → Whiteboard 0.1. Free, MIT, runs locally; accept the 0.x rough edges.
- Immutable-Linux distro or dev-workflow nerd → watch Toolpak. Nothing to install yet, but the proposal will shape how developer tools work on image-based desktops — contribute now.
The closing honesty note: Toolpak is a proposal with a funded prototype, not something you can install today, and Whiteboard is at 0.1.x by its own versioning. Everything else here is real, fetched, and live. My take: the interesting edge of the ecosystem is wherever the model, the robot, or the review canvas runs within arm’s reach. Of all the best new AI tools September 2026 put on the table, which would you deploy first?
References and further reading
- Aikido’s announcement of Altar-1 — compression methodology, CVE benchmark results, and deployment notes
- AikidoSec/altar-1 on Hugging Face — model card and serving commands
- Z.AI GLM-5.3 repository on Hugging Face — parent model and license text
- Liquid AI’s LFM2.5-2.6B release post — post-training pipeline and benchmark tables
- LiquidAI/LFM2.5-2.6B on Hugging Face — post-trained model weights
- LiquidAI/LFM2.5-2.6B-Base on Hugging Face — base model weights
- Feather Robotics — feather.dev — hardware specs, SDK docs, and pricing table
- TechCrunch’s profile of Feather Robotics — company background and funding
- devdotfast/whiteboard on GitHub — source code and project README
- install.dev.fast — Whiteboard desktop app downloads
- GNOME blogs — alatiera’s Toolpak design proposal
- #gnome-os:gnome.org on Matrix — feedback channel for the Toolpak proposal
Please let us know if you enjoyed this blog post. Share it with others to spread the knowledge! If you believe any images in this post infringe your copyright, please contact us promptly so we can remove them.