DEV Community: Rigor

The most overlooked AI story of 2026

Rigor — Fri, 04 Sep 2026 14:17:15 +0000

Ask around about the big AI stories of 2026 and you'll get a predictable shortlist. Meta abandoning open weights, then sort of walking it back. The OpenClaw saga that broke GitHub. Jensen Huang making his first post on X, ever. All legitimate picks. All heavily covered.

My answer is none of those. The most overlooked event of the year started in late August and is running as you read this: the Marin project out of Stanford is pretraining a 535-billion-parameter model with every metric streaming to the public in real time. It's the largest fully open, live-streamed frontier training run anyone has ever attempted. Total mainstream coverage so far: a few niche blog posts and one line in a newsletter digest.

Big and invisible at the same time. That's rare enough to be worth a closer look.

So what is this thing

Marin got started at Stanford in 2024, founded by David Hall and Percy Liang. The engineering work now happens under a nonprofit called Open Athena, with compute from Google's TPU Research Cloud and backing from the Huang Foundation, the Siegel Family Foundation, and Schmidt Sciences. Since around August 24, 2026, the project has been pretraining a frontier-scale model, and doing it in a way nothing at this size has been done before.

I want to be precise about "open" here. Every experiment begins as a GitHub issue stating its hypothesis up front. Configurations land as pull requests that outsiders can review while the run is in progress. Training metrics go public the moment a run starts. Data, code, recipe, final weights: all of it stays open afterward (Aoyii). Liang announced it himself, with typical understatement: "Marin 535B-A23B started training this week! As usual, the whole process is open" (X). His backers go further and call it "the largest open and live training run in history" (Open Athena) — and for once that's not puffery.

The numbers, with a small honesty note

The model is a 535.3-billion-parameter mixture-of-experts with 22.76 billion parameters active per token. The name just rounds those two figures. Exact specs live in the tracking issue.

Token budget comes with a footnote. You'll find two official figures: 18.75 trillion tokens in Liang's announced plan (80% pretraining, 20% midtraining), but 18 trillion in the repo config and the preregistration. The homepage's own target reads "Paloma eval loss, preregistered at 18T tokens: 2.04" (marin.community). Independent write-ups have flagged the same split (LamjinLab; Aoyii). Both numbers trace back to the project itself, so treat this as a loose end rather than a scandal. It's also the kind of discrepancy that only surfaces when you do everything in public, which is kind of the point.

Hardware: eleven NVIDIA GB200 NVL72 racks, roughly 792 GPUs, rented from CoreWeave. About 2.7×10²⁴ FLOPs over a ~100-day trajectory, projected to wrap around December 1, 2026 (KAD8; 36kr). As of Liang's early-September update, the run was 13% complete (X).

Follow the money, because it leads somewhere awkward

The compute is funded by the Jen-Hsun and Lori Huang Foundation. Yes, that Huang. Liang credited them directly: "This hero run would not be possible without the generous support of the Jen-Hsun and Lori Huang Foundation, which provided the funding for the compute (Coreweave)" (X).

There's independent corroboration, sort of. A regulatory filing reported by Reuters says the foundation has purchased roughly $108.3 million of CoreWeave computing capacity for academic and nonprofit AI and scientific research (Yahoo Finance). Two caveats: the filing predates the run and never names Marin, so the link to this specific run rests on Liang's public credit. And the mechanism matters — it's a capacity purchase (unlikely a hardware gift).

There's also a conflict-of-interest thread worth pulling. NVIDIA itself invested $2 billion in CoreWeave and signed a $6.3 billion capacity guarantee, so the money behind this run sits close to NVIDIA's own commercial interests. The blunt take in the coverage, that this "isn't purely philanthropic," seems fair (AI Weekly). You can hold that thought and still note that the compute is real and the run is genuinely open. Both things are true at once.

They did the homework in public too

A 535B run sounds reckless for a nonprofit, but the groundwork was done the same way as everything else. Before the hero run, the team climbed a four-rung scaling ladder from 1.6B-A61M (48B tokens) up to 27.7B-A1.2B (926B tokens), burning about 1% of the final compute budget. On top of that sits Delphi, their scaling suite, which forecasts pretraining outcomes to within 0.2% error (RuntimeWire).

The openness has already paid off once, too. Early in the run, the community spotted a loss spike near training step 500 and traced it to router_bias in real time, straight from the public metrics (KAD8). At a closed lab, that's an internal incident report nobody outside ever hears about. Here it was a public debugging session.

The announcement also traveled. Liang's post passed 800,000 views and drew an endorsement from Andrew Ng, who called it "a precious demonstration" of openness and noted that public release of AI research "used to be the industry norm" (Aoyii; 36kr). One flag before you quote him: those lines circulate via relay pages rather than anything first-hand.

Why nobody covered it

Here's my theory, and I keep coming back to it. Every other major 2026 story is a release. A model ships, benchmarks get quoted, articles get written. That's the template.

Marin isn't a release. It moves the frontier from open product to open process, and release-shaped coverage has no slot to put it in. So it landed in niche press, got one digest mention (The Neuron), and vanished from the conversation.

The sharpest proof of the blind spot comes from inside Stanford itself. HAI published an op-ed on August 4, 2026 arguing that open-weight models aren't enough and the field needs truly open source AI, citing DeepSeek and Qwen (Stanford HAI). It did not mention Marin, which was weeks from starting the largest truly open training run ever attempted, down the hall, so to speak.

Marin is mid-run — 13% as of early September, with December 1 a projection rather than a promise. The tracking issue notes that up to ~25% token-budget degradation would be handled by shrinking the token horizon and retiming the learning-rate decay, and even friendly coverage calls the schedule "an operating plan rather than an immovable contract." Wait for the weights before declaring victory.

On the money: confirmed at the program level, inferred at the run level. The Huang Foundation's $108.3M of CoreWeave capacity is real per the filing; the tie to this run rests on Liang's statement.

Two source-quality notes. KAD8 calls Liang a "Simile AI founder," which is garbled — he's a Stanford professor and Marin's co-creator alongside David Hall, and per his ICLR bio also a co-founder of Together AI and Simile AI. And the Andrew Ng quotes should be attributed carefully, since they only circulate through relay pages.

Also, OpenEuroLLM's July 31 first-models milestone was scheduled, but nothing I found confirms a release actually happened (OpenEuroLLM). Unshipped until proven otherwise.

Why this is my pick

If I only get to name one overlooked story for 2026, it's Marin, and the reasoning is simple. OLMo 3 set the openness bar in November 2025 (Ai2). Marin is the first 2026-native attempt to carry that ethos to frontier scale, and it says so itself, crediting BLOOM, Pythia, LLM360, and OLMo as lineage. The smaller runs beneath it tell the same story: Marin 8B was the first fully open JAX model back in July 2025, a 32B followed that October, and the project landed an ICLR 2026 invited talk.

It's the only 2026 event that changes what "open" will mean in 2027, rather than just participating in the current meaning. And it's happening on roughly 792 GPUs in full public view, every metric there for anyone curious enough to look, while almost nobody looks.

That's what makes it the answer. The most overlooked story isn't the one that got buried. It's the one running in plain sight.

If you're following the run yourself, I'd genuinely like to hear what you're seeing in the curves!

Underrated AI Coding Agent Harnesses in 2026: pi, mini-swe-agent, Crush, Plandex, Amp and More

Rigor — Thu, 03 Sep 2026 08:48:59 +0000

If you have only heard of Claude Code, Codex CLI, or Gemini CLI, you are missing a big part of the picture. A handful of smaller, lesser-known coding agent harnesses are quietly matching or beating the big names on benchmarks. The standout is pi, and there is a solid supporting cast behind it: mini-swe-agent, gptme, Crush, Plandex, and Amp, plus some deeper cuts worth knowing about.

Why the harness matters more than the model

A useful definition comes from an independently maintained catalog of over 100 harnesses that gets rescored weekly: a model answers, an agent acts, and the harness is the runtime that turns one into the other. The same catalog found that swapping the harness around a model can move its pass rate more than upgrading the model itself does. One example: the identical model scored 23% versus 52% on SWE-bench Pro depending only on which harness wrapped it.

Independent leaderboards back this up. Claude Opus scores 77% on Terminal-Bench inside Claude Code but 93% inside Cursor, according to a comparison called "The Harness Effect." The official Terminal-Bench leaderboard shows the same underlying models spread from 51.82% for Claude Code with Opus 5 down to 20.30% for xAI's Grok Build, with Codex CLI landing in between at 37.27%. And the actual top of the Terminal-Bench 2.0 board was not a household name at all: a research harness called "vix" running Claude Opus 4.7 hit 90.2%, ahead of Codex CLI's 82.2% with GPT-5.5. Obscure clearly does not mean worse here.

The lesser known but genuinely excellent ones

pi is the minimal, hackable option, and it is the one people keep bringing up. It was built by Mario Zechner, known for creating libGDX and RoboVM, who joined Armin Ronacher's company Earendil in 2026. Ronacher is the creator of Flask and Jinja. The project moved from Zechner's personal badlogic/pi-mono repository to earendil-works/pi in May 2026, with the core staying MIT licensed even as some future enterprise features may shift to a Fair Source or proprietary model.

The whole idea behind pi is "adapt pi to your workflows, not the other way around." It ships with a small core, just the agent loop, file tools, bash, and model providers, and everything else comes through TypeScript extensions, skills, prompt templates, and themes that get shared as packages via npm or git. There is no sub-agent system and no plan mode built in. If you want a feature, you ask pi to build it for you.

Zechner has been vocal about why this matters to him: exactly controlling what goes into the model's context yields better outputs, while most existing harnesses inject things behind your back. He has described Claude Code as a spaceship packed with 80 percent functionality he never uses. And this restraint is not just philosophy for its own sake. With nothing but read, write, edit, and bash tools, no web search, no compaction, and a system prompt under 1,000 tokens, pi took 8th place on Terminal-Bench running Opus 4.5.

A few practical details worth knowing if you are considering it: it routes to Anthropic, OpenAI, Google, and more than 20 other providers using your own API keys. People run local GGUF models like Qwen through llama.cpp inside it. It works on Windows through Git Bash in Windows Terminal. Hooks are stateful TypeScript modules with UI helpers built in. It is also the engine powering OpenClaw. Star counts quoted across different sources range widely, from around 61,000 up to roughly 101,000, depending on when the snapshot was taken. The trade-off is real: you get more control and better token efficiency, but you end up assembling more of your own workflow.

mini-swe-agent is the radically simple, research-grade choice. It comes from the Princeton and Stanford team behind SWE-bench and SWE-agent, and the whole thing is genuinely about 100 lines of code. Despite that, it scores over 74 percent on SWE-bench Verified, is MIT licensed, and has around 6,900 stars. It is best for people who want an agent they can read and audit in an afternoon, and it doubles nicely as a research baseline for anyone studying how these systems work.

gptme is the local-first veteran of the group. It was one of the very first agent CLIs, dating back to spring 2023, and it still runs anywhere a terminal does, whether that is your laptop, an SSH session, tmux, a headless server, or a CI pipeline. It is provider-agnostic, comes with shell, Python, and vision tools plus a plugin system, and supports fully local operation through llama.cpp. Its community is smaller, around 4,400 stars, but it has stayed active and relevant for years, which says something.

Crush is the best-looking terminal agent you will find. Built by Charmbracelet under the tagline "glamorous agentic coding," it has around 27,900 stars. It works with any LLM through OpenAI- or Anthropic-compatible APIs, lets you switch models mid-session without losing context, integrates with LSP, extends through MCP, and runs cleanly across macOS, Linux, Windows, and the BSDs. It is technically Charm's fork of the original OpenCode, released under the FSL-1.1-MIT license.

Plandex is built for big, multi-step tasks. It is a terminal tool designed to plan and execute large coding jobs that span many steps and dozens of files, and it keeps a cumulative diff review sandbox so AI-made changes stay separate from your actual files until you approve them. It has around 15,600 stars.

Amp is the polished, opinionated commercial option. Sourcegraph launched it in May 2025, and it spun out as its own independent company, Amp Inc, in December 2025 under co-founder Quinn Slack. It gives you one agent across web, Mac, iPhone, iPad, and terminal, along with a feature called Orbs, which are per-thread cloud machines that let you send a prompt, close your laptop, and have the agent keep working. The team's stated editorial approach is blunt: if they do not use and love a feature, they kill it. It is well liked among people who do heavy agentic coding work, though it is not open source.

The deeper cuts

A few more names worth knowing if you like exploring the edges of this space. Tau, built by Hugging Face, has around 2,600 stars and is a small, readable Python terminal agent inspired by pi that doubles as a teaching codebase for understanding how harnesses work under the hood, released under MIT. hax has around 700 stars and is a minimalist, terminal-native agent written in C that runs in just a few megabytes of RAM, with strong llama.cpp support. Paseo has around 15,600 stars and is a no-telemetry parallel runner that can drive Claude Code, Codex, Copilot, OpenCode, and pi agents side by side. Orca, from Stably, has around 58,000 stars and runs a fleet of parallel agents, each working inside its own git worktree.

The familiar names, for reference

To put all of this in context, it helps to know where the popular options stand. opencode is the most-starred open-source agent out there, with around 203,000 stars, MIT licensed, and provider-agnostic. Goose started at Block and now lives under the Linux Foundation's Agentic AI Foundation, Apache-2.0 licensed, with around 53,900 stars and support for more than 70 MCP extensions. Aider remains the classic diff-based pair programmer, with around 48,600 stars. OpenHands sits around 85,600 stars, and Cline, which is VS Code native with Plan and Act modes plus MCP support, has around 67,000.

Then there is the first-party trio everyone already knows: Claude Code, which tops the Terminal-Bench 4.0 board at 51.82 percent when paired with Opus 5, Codex CLI, and Gemini CLI. They anchor the benchmark tables and are worth knowing, but at this point they are not really the interesting edge of the design space anymore.

How to actually choose one

If you want control, transparency, and token efficiency, go with pi, or mini-swe-agent if you would rather read the entire agent in one sitting.
If local models and privacy matter most, look at gptme or hax for their llama.cpp-first approach, or Goose paired with local providers.
If you want a batteries-included open source option that works with any provider, opencode is the safe pick, with Crush as the prettier alternative.
If you are tackling big multi-file tasks and want reviewable diffs, Plandex is built exactly for that.
If you want something polished and commercial with team threads and background cloud agents, Amp fits the bill.
If you need to run parallel fleets across many tasks at once, layer Orca or Paseo on top of whichever harness you already use.

A few caveats worth keeping in mind before you commit to any of these. Star counts and benchmark scores are point-in-time snapshots, and this space shifts on a monthly basis. Some Terminal-Bench links now point to the current 4.0 board rather than the historical 2.0 results referenced above. Amp's documentation does not list pricing. And pi's 8th place finish on Terminal-Bench was a single full run reported by its own author, so treat it as a strong signal rather than an independently verified result.

A Field Report on Open-Weight AI in 2026

Rigor — Wed, 02 Sep 2026 14:17:26 +0000

If you last checked in on open-weight models a year ago, the picture has inverted. The clearest takeaway from the first half of 2026 is that choosing an open model is no longer a single-model decision. It is a portfolio decision, made across a crowded field of frontier-class releases documented in Digital Applied's H1 2026 retrospective. Here is what actually shipped, who shipped it, and what "open" turned out to mean.

Ten weeks, five Chinese labs

Between April and June 2026, five Chinese labs shipped open-weight frontier models back to back: DeepSeek V4, Kimi K2.6, GLM-5.1, Qwen 3.6, and MiniMax M3, according to Sunset Browser's roundup. April alone was absurd across the whole industry. DeepSeek V4, Qwen 3.5-Omni, Gemma 4, Meta Muse Spark, GPT-6, and Claude Opus 4.7 all landed within a single calendar month (Fazm).

DeepSeek V4 set the pace

DeepSeek V4 shipped on 24 April 2026 under an MIT licence in two tiers. V4-Pro sits at 1.6T parameters with 49B active; V4-Flash at 284B with 13B active. Both carry a 1M-token context window and 384K maximum output (DataMy). Flash is the one you can actually run: a Mixture-of-Experts build aimed at coding, tool use, and agentic workflows, downloadable through LM Studio. Moonshot's Kimi K3 is similarly a download away on Hugging Face, deployable with a standard vLLM server.

Four labs, four meanings of "open"

Openness itself is uneven. Tracking four flagship open-weight releases since mid-June, the lag between a model appearing on a public API and its weights being published ran zero days for DeepSeek V4 Flash and GLM-5.2, and eleven days for Moonshot's Kimi K3 (Alephant). Alibaba has institutionalised the split: since 2026 it runs a deliberate two-track strategy, open weights for the community and on-premise deployment, with proprietary Plus and Max tiers sold through Alibaba Cloud (Tongyis). Even Thinking Machines joined the movement, releasing its Inkling weights openly on Hugging Face. That is a real shift in transparency and licensing, though it comes with notable restrictions (Checkingmarket).

Google came back, and everyone else kept coming

Google DeepMind's Gemma 4 arrived as a four-model open-weight family combining Thinking Mode reasoning, native multimodality, 256K context windows, and Apache 2.0 licensing. It benchmarked 89.2% on AIME 2026 with an 86.4% score on a second benchmark (LinkedIn).

Zhipu's GLM line kept the pressure on. GLM-5.2, released on 13 June 2026, brought MIT-licensed weights, a 1M-token context window, and top-of-open-source coding scores (Kie.ai). Its successor GLM-5.3 Flash is flagged as one of the most important open-weight releases to watch in late 2026, though reviewers caution it is not an uncensored chat model and not the easiest local model to fit on a consumer GPU (AirMore).

Elsewhere, the releases kept stacking up. Tencent's Hy4 preview landed on 28 August 2026 as an open-weight model you can download and run on your own hardware (LLM Stats). Nvidia shipped Nemotron 3.5 Lightning on 11 August 2026, a free open-weight model aimed at autonomous agent workloads (Andrew.ooo). Mistral confirmed a sparse Mixture-of-Experts family entering early access in July 2026, its boldest open-weight bet yet, backed by a €4bn EU data-centre buildout (AI TechConnect).

Switzerland's Apertus got a more sober verdict one year on: "When it comes to agentic capabilities, despite considerable progress, Apertus is not yet at the level of other open-weight models," Binaghi told SwissInfo.

The policy fight went public

On 27 July 2026, Dario Amodei published a direct rebuttal to industry chatter suggesting Anthropic secretly wants open models banned. The clarification reshapes how founders, policymakers, and builders should read the open-weights debate heading into late 2026 (Kalinga.ai).

Choosing between them

The comparison infrastructure has matured alongside the models. BenchLM maintains direct comparison tables across GPT-5, Claude, Gemini, DeepSeek, Llama, and dozens of other frontier and open models. HypeBench runs a comparison of DeepSeek V4 Pro, GLM-5.3, Qwen 3.8 Max, and Kimi K3 with workload tests and access checks. Artificial Analysis computes its Intelligence Index from output tokens per task divided by speed, weighted across benchmarks.

Older standbys still hold their niches. Qwen 3.5 leads most open-source benchmarks under Apache 2.0 licensing, DeepSeek R1 dominates math and reasoning with a 97.3 on AIME, and Llama 3.3 70B remains the safe, well-supported generalist (Toolhalla). A Turkish roundup ranks Llama 4, Qwen 3.6, and DeepSeek V4 by benchmarks, hardware requirements, and true cost (GetAIPerks). Capability maturity is uneven down the stack: Mistral Large 2 and Qwen 2.5 call tools natively, Llama 3.1 weakly, DeepSeek V3 at a mid level (Omeronal).

The loose ends

Meta is the big one. Muse Spark shipped inside that crowded April window, but nothing in the coverage here establishes whether Meta's 2026 models are open-weight at all. Until that question gets an answer, Llama's open-source standing rests on older releases like Llama 4 and 3.3.

The second loose end is smaller but worth flagging. A few headline numbers arrived without their backstory: which benchmark sits behind Gemma 4's 86.4% score, and what the full weight-lag figures look like for the four-release tracker beyond the three models named. A number that cannot say where it came from cannot tell you much.

The quiet side of open

There is a whole side of this story the headlines skip, and it falls into four clusters: the genuinely open camp (IBM Granite 4.x, Hugging Face's SmolLM3, Ai2's OLMo line, Stanford's Marin, OpenEuroLLM, Sarvam, plus GLM-5.3's full-weights release), open multimodal, video, and robotics models, two major Chinese labs (Baidu and ByteDance), and the open-weights policy track. Two of those blind spots, IBM's Granite and Hugging Face's SmolLM3, have documented substance behind them, and both get a closer look below.

Granite 4.x, the omission with the most substance

Granite is the omission with the most substance behind it, and its release profile looks nothing like the hype cycle. Granite 4.0 runs a hybrid Mamba and transformer architecture and ships with signed model weights and documented training data, positioned explicitly for sensitive use in healthcare and the public sector (Geeky Gadgets). Granite 4.1 is the Apache 2.0 dense family in 3B, 8B, and 30B sizes, and the 3B runs from roughly 2 GB of memory at Q4 quantization, with hardware-fit guidance for each size (The AI Bench). Context stretches to 131k across the generation: the 4.1 8B supports up to 131,072 tokens (free.ai), and the 4.1 30B sits on Artificial Analysis' leaderboard of 250+ models at 131k (Artificial Analysis). Granite 4.2 8B is IBM's dense reasoning model, served on OpenRouter at $0.10 per million input tokens and $0.15 per million output tokens (OpenRouter). The 1B and 3B variants are IBM's first mixture-of-experts Granite models, designed for low-latency use (Ollama), with a lineage that runs back through Granite 3 MoE, which activated about 3B of a 10 to 15B total per token (RunLocalAI).

Want to run it yourself? Granite runs locally through Ollama on modest hardware (LinkedIn), a free-tier granite-4.0-micro endpoint is available through a single OpenAI-compatible API (UnoRouter), and the brand stretches into small specialists: a 258M-parameter Granite Docling model with a hosted demo (Free2AI Tools) and a 278M multilingual embedding model open-sourced on GitHub (Toolify).

A permissive licence, a hybrid MoE family, signed weights, and documented data. That is the release profile regulated industries and the public sector actually ask for, and it is a different story than the weights-only headline labs.

SmolLM3, small on purpose

The other documented blind spot is small on purpose. Hugging Face's Smol Models team shipped SmolLM3, a compact open-weight 3B model built around efficiency, multilingual reach, and long-context reasoning (Pure Neo). It carries a 128,000-token context window and dual reasoning modes, and it is reported to outperform larger models at that footprint (CTOL, ScaleByTech). Where it runs is the point: local execution (ML Hive) and personal devices (Techzine), not someone else's cloud. Its predecessor SmolLM2 was trained on 11 trillion tokens, including custom math, code, and instruction datasets (Analytics India Magazine).

What makes it fully open is the training set. SmolLM-Corpus is itself published, combining Cosmopedia v2 (quality synthetic educational content) and FineWeb-Edu among its three main components (Noze), and other locally-runnable models are now being trained on it too (The Unwind AI).

For everything else, 2026's story is already clear: open weights stopped being the alternative and became the market.