Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

The open source LLM fine tuning stack is being rebuilt around agents, not models

On 30 September, Cloudflare open sourced Forge, a generation pipeline that turns an API spec into SDKs, CLIs, docs and MCP servers. NVIDIA also published Kumo Tabular, a foundation model that predicts from tables in one forward pass, with no training or feature engineering. Neither is a fine tuning tool, but both show where the open stack is moving.

AI & modelsAnalysisRachel NwosuPublished: 30 September 20269 min readSources 15
The open source LLM fine tuning stack is being rebuilt around agents, not models
On 30 September, Cloudflare open sourced Forge, a generation pipeline built to produce SDKs, CLIs, docs and MC

The newest item here is not a fine tuning framework. On 30 September, Cloudflare published Forge, an open source, pluggable generation pipeline. It turns an API specification into SDKs, CLIs, docs and MCP servers. The company's blog says Forge already generates the output required for the cf CLI. It will power Cloudflare's API documentation and SDKs over the next few months.

The stated reason is not developer convenience. Cloudflare says it built Forge "because we needed it ourselves in order to treat agents as our customers." The API behind it carries more than 3,500 operations. They span services written in Rust, Go, TypeScript and Python. That framing is the tell. The tooling around a model, the glue that turns weights into a product, is now what gets open sourced and standardised. Often it comes before the models themselves.

NVIDIA made the same move on the same day, in a different corner. Its Kumo Tabular entry on Hugging Face, published 29 September, describes an open foundation model for tabular classification and regression.

Given labelled rows, it returns class probabilities or numbers in a single forward pass. It needs no training, no tuning and no feature engineering. It ships in three sizes from 28M to 215M parameters and was pretrained only on artificial data. It is released under the OpenMDW-1.1 licence for commercial use. NVIDIA says it ranks first on TabArena, BeyondArena, TALENT and ScoringBench. The pitch is explicitly a replacement for the retrain-and-deploy cycle. Gradient-boosted trees imposed that cycle on enterprise machine learning for two decades. So the open stack is splitting in two. One branch removes the fine tuning step from whole categories of work. The other branch still needs weights changed, and its surrounding machinery is being opened up.

The cheap-inference branch is growing fast

Fermion Research released Phonon-2 on 30 September, an English speech recognition model. Its encoder stores every weight as one of five learned levels at about 2.1 bits per weight. The download is 164 MB. According to the company's research page, it averages 5.21 percent word error across the Open ASR Leaderboard's seven English sets. Every open model that scores better is at least 5.8 times its size. It turns an hour of audio into text in roughly 20 seconds on a MacBook Air. The weights are CC-BY-4.0, derived from NVIDIA's Parakeet TDT 0.6B v3.

The pattern repeats with Meta's OpenZL, which shipped v0.3.0 on 29 September. The release notes describe a new LZ engine using PivCo Huffman as its entropy backend. Decompression is 33 percent faster than v0.2.0 and 144 percent faster than Zstandard at equivalent settings.

More interesting for anyone training or serving models: the Compression Transformer. It is a small neural network that builds numeric compression graphs on the fly, per input, with no per-source training. Meta evaluated it on 868 families of numeric streams, 34,737 files and 17.9 GB. Meta says it compresses 35 percent better on average than zstd -19. It also lands within 1.1 percent of graphs that were trained. Smaller weights, better compression, fewer tokens spent. That is the direction of travel, and it costs the fine tuning industry work.

TypeSafe's Jev is the clearest example of a model built to need none of the usual adaptation. It answers Choice, Score and Noul questions rather than generating text. That makes it cheap enough that the interesting problem becomes routing, not training.

The Register reported on 29 September that Jevstiller, an open source project, distills Jev into a local student model. It aims for 98 percent agreement with the hosted service. It also answers familiar requests on a device CPU in as little as 15 ms. The project's own documentation, quoted by The Register, says confident inputs are answered locally and everything else goes upstream. That is fine tuning by another name, and it happens on the user's machine, against a teacher the user does not control.

What a fine tuning stack now has to defend

OpenAPPA, pushed to GitHub on 30 September, sits between an agent and its tools. Before every call it answers one question: is this data allowed to go to this destination? It tracks the sensitivity and trust of everything an agent reads. It enforces policy written in declarative TOML.

The project's benchmarks claim no scored attack succeeded in 1,320 evaluations while completing 88 to 90 percent of tasks. They compare this with Microsoft FIDES letting 28 to 35 percent of attacks through. These are the authors' own numbers, run on their own suites. The UAI protocol, also published 30 September, attacks the same gap from the identity side. An agent gets an identity bound to an accountable owner and a time and jurisdiction bounded passport. It also gets signed, hash-chained attestations of what it did.

The README is careful in a way most projects are not. "UAI does not claim that an AI agent is safe," it says. "It provides a mechanism for making an agent's actions attributable and independently verifiable." Both projects exist because the alternative has visibly failed this year. That alternative is trusting the system that performed an action to also write the log.

The Guardian reported on 30 September that the FTC has opened an industry-wide investigation into Anthropic, OpenAI and others. It is the first official US enforcement action touching rogue AI agents, with formal demands for information and compelled testimony planned. The New York Post first reported the news, according to The Guardian.

Ars Technica reported the same day that OpenAI has pushed its IPO back. Sam Altman told reporters at the company's developer day that it would not "barrel all guns blazing towards an IPO" while it still needs to "make confident safety decisions." Ars put the company's valuation at $852 billion. It said the company is in talks to raise $30 billion or more at about $1.4 trillion. A non-profit called Legal Advocates for Safe Science & Technology filed suit in California the same day. It seeks better evaluation, monitoring and training practices.

"Given what we see in terms of progress and development [in AI], the frequency and sophistication of these hacking instances are only going to increase," said Vivian Dong, programs director at LASST, in comments reported by Ars Technica.

Training practices are now a legal exposure. That changes what a fine tuning stack is for.

The governance layer is being open sourced too

OpenClaw Enterprise, announced 30 September, is the most direct attempt to answer this. It is an open source, vendor neutral control plane for persistent agents. It offers multi-tenancy, hard boundaries between trusted and untrusted workloads, sandboxing, LLM-based reviews and fine-grained permissions. The blog says the project started at OpenAI and was donated to the OpenClaw Foundation. It is now developed with Red Hat and NVIDIA. It is pre-1.0 and described as suitable for internal pilot workloads. OpenAI is already running OpenClaw agents with full access to codebases and plugins, per the announcement.

The UK AI Security Institute's Inspect, documented on 30 September, covers the measurement side. It is a framework for frontier evaluations with composable datasets, solvers and scorers. It has more than 200 pre-built evaluations. It supports running arbitrary external agents including Claude Code, Codex CLI and Gemini CLI, with sandboxing in Docker, Kubernetes or Modal.

Then there is the quieter, more practical layer. A GitHub repository published on 30 September catalogues known failures in ChatGPT Work and Codex after OpenAI's 29 September DevDay. It links to user reports rather than confirmed bugs. Its central complaint: long agent runs fail and the consumed usage is not refunded, because there is no checkpoint to resume from. One cited Codex desktop report says a repair run burned about 50 percent of a weekly allowance after 7 hours 15 minutes. OpenAI did not confirm any of it. The repository's author says GPT-5.6 Sol performs well through OpenCode, which points at infrastructure rather than models.

Vincent Schmalbach, writing on 30 September, makes the commercial version of the same argument. He says OpenAI gave him $5,000 in usage credits across two Codex accounts. He says it also told him the allowance on his $200 subscriptions would be halved from 30 October, when Pro 200 moves from 20x to 10x the Plus allowance. He plans to downgrade to Plus and spend the difference on Claude subscriptions. That is one customer's account, not a market survey. But the direction is consistent with a stack where the model is a commodity and the harness is the product.

The missing piece

The topic that sent me looking, an open source LLM fine tuning stack, is the thinnest part of this article. Two entries come close. Bespoke Labs' Nimble, pushed 30 September, publishes data, a training recipe and a 9B model that makes typed decisions. The README is explicit that it did not distill from Jev. The point is to show how to curate data, train and serve, and to encourage more research.

OpenResearch, also published 30 September, is a local-first desktop and CLI workspace. It gives coding agents literature search across alphaXiv, bioRxiv and PubMed. It also gives them experiment execution on Slurm, Kubernetes, Modal or Hugging Face Jobs. And it keeps a local SQL record of every run so agents can learn from past results. Neither is a trainer. Both assume the model already exists and the hard part is orchestration, data provenance and evaluation. That is a real shift, and it is not obviously good news for anyone whose product was a wrapper around a training script.

One counterweight, from Rest of World on 30 September: China's domestic open model hubs are scaling hard. Alibaba's ModelScope, launched in 2022, now hosts more than 170,000 models. OSChina's MoArk, launched in 2023, serves some 20,000.

OSChina chief executive Xu Yong told the outlet that not everyone can use a VPN all the time. He said China is developing an independent ecosystem faster in AI than it did in the internet era. Rest of World also reports that Nvidia announced in September it was acquiring Hugging Face for $12.9 billion. Sovereign stacks mean sovereign fine tuning, and that work will not happen on the same tooling as everyone else's.

The week's releases share a premise: stop training when you can route, compress or verify instead. Whether that holds for the workloads that genuinely need new weights is the question none of these announcements answered.

Comments 0

Sources

15
  1. 01Introducing Forge: the open source pipeline for generating SDKs, CLIs, docs, and moreEN
  2. 02NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular PredictionEN
  3. 03Phonon-2: most accurate open speech recognition model in a 164 MB downloadEN
  4. 04Release OpenZL v0.3.0EN
  5. 05Open source tool distills Jev so you can run it locallyEN
  6. 06OpenAPPA: Deterministic guardrails that don't break agentsEN
  7. 07UAI: An open protocol for identity and accountability of AI agentsEN
  8. 08US trade regulator opens investigation into AI giants including Anthropic and OpenAIEN
  9. 09OpenAI delays IPO over AI safety concernsEN
  10. 10OpenClaw Enterprise - The Open Agent PlatformEN
  11. 11Inspect: An open-source framework for large language model evaluationsEN
  12. 12OpenAI after DevDay 2026: known issues for serious development workEN
  13. 13OpenAI Is Paying Me to Move My Business to AnthropicEN
  14. 14Nimble: Data, Model, Recipe for an Open Jev (From Bespoke Labs)EN
  15. 15The open-source AI platforms vying to become China's Hugging FaceEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.