# forck.live — 11–17 August 2025

> AI lab announcements, news reports and podcasts.

forck.live follows what companies announce about their own products. Press may
corroborate an official story; it never leads and never appears as an item of its
own. A missing date stays missing. A title is the publisher's words — where a
company does not publish in English, an English rendering is listed separately
as `Title in English` and named as this site's translation, never as theirs.

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2025-W33, 11–17 August 2025, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

Every entry is one first-party announcement, and carries the same eight fields
in this order before any others:

    Company · Announced · Category · Coverage · Announcement · Group · Source · Record

`Source` is the company's own page the announcement was traced to and `Record`
is this site's page for it. Both are safe to cite.

`Coverage` is a number of outlets, or `not counted` — which means this record
reached no count of its own, never that nobody wrote about it.

`Category` is this tracker's own analysis, and is authoritative wherever it is
stated. `not stated` means either that nothing has analysed the post yet or
that the analysis placed it outside this taxonomy; neither is a verdict about
the announcement.

`Announcement` is whether the post is news of something the company did, rather
than a how-to, a customer story, a changelog listing or a staffing note.

`Group` is which of this record's groups the entry falls in, carried on the
entry itself so it holds where the groups are headings and where they are not.
One of: `models`, `covered`, `announcements`, `routine`.

Anything after `Record` is held detail that only some announcements carry.

- This week's page: https://forck.live/week/2025-W33
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### Supabase — The Vibe Coder's Guide to Supabase Environments

- Company: Supabase (supabase.com)
- Announced: 2025-08-17T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/the-vibe-coders-guide-to-supabase-environments
- Record: https://forck.live/items/21336-the-vibe-coder-s-guide-to-supabase-environments
- Subject: Supabase

Setting up separate development and production environments does not have to be painful. This guide shows you how to build a professional deployment workflow for your Supabase project.

### Supabase — Vibe Coding: Best Practices for Prompting

- Company: Supabase (supabase.com)
- Announced: 2025-08-16T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/vibe-coding-best-practices-for-prompting
- Record: https://forck.live/items/21339-vibe-coding-best-practices-for-prompting
- Subject: Supabase

Supabase published a guide on prompting strategies for AI coding assistants, covering a three-layer prompt structure, iterative refinement, and templates for common tasks.

### Supabase — The Vibe Coding Master Checklist

- Company: Supabase (supabase.com)
- Announced: 2025-08-16T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/the-vibe-coding-master-checklist
- Record: https://forck.live/items/21338-the-vibe-coding-master-checklist
- Subject: Supabase

Supabase published a guide for developers using AI coding tools to prepare their applications for production. The guide covers security auditing, performance optimization, and deployment planning, with specific recommendations for using Supabase's integrated platform features like authentication, Row Level Security, and database management.

### Supabase — Testing for Vibe Coders: From Zero to Production Confidence

- Company: Supabase (supabase.com)
- Announced: 2025-08-16T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/testing-for-vibe-coders-from-zero-to-production-confidence
- Record: https://forck.live/items/21337-testing-for-vibe-coders-from-zero-to-production-confidence
- Subject: Supabase

Supabase published a guide for developers on building a testing strategy that prevents production issues, focusing on integration tests against a real local Supabase stack including Postgres, Auth, and Row Level Security.

### Replit — Introducing Image Generation for Replit Agent: Bring Your Ideas to Life in Pixels

- Company: Replit (replit.com)
- Announced: 2025-08-15T22:45:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/image-generation
- Record: https://forck.live/items/17945-introducing-image-generation-for-replit-agent-bring-your-ideas-to-life-in-pixels
- Subject: Replit Agent

Replit added AI-powered image generation to Replit Agent, allowing users to create and integrate images directly within their projects via chat prompts. The feature costs about $0.04 per image under Effort-Based pricing.

### Together AI — Fine-Tuning Small Open-Source LLMs to Outperform Large Closed-Source Models by 60% on Specialized Tasks

- Company: Together AI (together.ai)
- Announced: 2025-08-15
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/fine-tune-small-open-source-llms-outperform-closed-models
- Record: https://forck.live/items/2415-fine-tuning-small-open-source-llms-to-outperform-large-closed-source-models-by
- Subject: Inference platform

Together AI and Parsed demonstrate that fine-tuned small open-source models, using Together AI's fine-tuning platform, can outperform large closed-source models by 60% on specialized healthcare tasks while reducing inference costs by 10-100x.

### Lovable — Lovable Goes Back to School

- Company: Lovable (lovable.dev)
- Announced: 2025-08-14T19:02:00+00:00
- Category: pricing-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/back-to-school-discount
- Record: https://forck.live/items/12135-lovable-goes-back-to-school
- Subject: Lovable / AI app builder

Lovable is offering a 50% discount on its Pro plan for students globally, verified via school email or student ID. The company is also organizing Back-to-School Hackathons in partnership with power users.

### Hugging Face — Kimina-Prover-RL

- Company: Hugging Face (huggingface.co)
- Announced: 2025-08-14T12:13:01+00:00
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/AI-MO/kimina-prover-rl
- Record: https://forck.live/items/1635-kimina-prover-rl
- Subject: Platform

AI-MO released an open-source training pipeline kimina-prover-rl for Lean 4 theorem proving and two open-weight models: a 1.7B-parameter model scoring 76.63% Pass@32 on MiniF2F and a 0.6B-parameter model scoring 71.30% Pass@32, each setting a new state of the art for open-source models in their size.

### Genspark — Introducing Genspark AI Developer - The L4 Autonomous Coding Agent That Lets Everyone Build

- Company: Genspark (genspark.ai)
- Announced: 2025-08-14
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.genspark.ai/blog/genspark-ai-developer
- Record: https://forck.live/items/18271-introducing-genspark-ai-developer-the-l4-autonomous-coding-agent-that-lets
- Subject: Genspark

### Baidu — FastDeploy 2.0: A Large-Scale Model Inference and Deployment Toolkit with Native Support for ERNIE 4.5

- Company: Baidu (baidu.com)
- Announced: 2025-08-14
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ernie.baidu.com/blog/posts/fastdeploy2.0/
- Record: https://forck.live/items/4654-fastdeploy-2-0-a-large-scale-model-inference-and-deployment-toolkit-with
- Subject: ERNIE

Baidu announced FastDeploy 2.0, an end-to-end toolkit for efficient deployment and high-performance inference of large models, with native support for the ERNIE 4.5 family. It features high-throughput inference using Expert Parallel and Prefill-Decode Disaggregation, near-lossless 2-bit quantization enabling trillion-parameter models on a single GPU, and optimizations for heterogeneous hardware. Benchmark results show significant throughput gains over vLLM for ERNIE and other models.

### Replit — Introducing App Storage – building apps with images, video, and PDFs just got easier

- Company: Replit (replit.com)
- Announced: 2025-08-13T17:01:07+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/app-storage
- Record: https://forck.live/items/17946-introducing-app-storage-building-apps-with-images-video-and-pdfs-just-got-easier
- Subject: Replit Agent

Replit announced App Storage, a built-in object storage solution for hosting and saving uploads like images, videos, and documents within applications. The feature, previously called object storage, was renamed to highlight Agent's new capability for seamless setup. App Storage includes per-user permissions, a database table for file metadata, and scaffolded UI flows.

### Hugging Face — Arm & ExecuTorch 0.7: Bringing Generative AI to the masses

- Company: Hugging Face (huggingface.co)
- Announced: 2025-08-13T14:55:10+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/Arm/executorch-0-dot-7
- Record: https://forck.live/items/1636-arm-executorch-0-7-bringing-generative-ai-to-the-masses
- Subject: Platform

Arm and ExecuTorch announce that KleidiAI will be enabled by default in ExecuTorch 0.7 beta, bringing automatic AI acceleration to Arm CPUs, including older devices with SDOT support, enabling on-device generative AI like Llama 3.2 on billions of devices.

### Vercel — How Coxwave delivers GenAI value faster with Vercel

- Company: Vercel (vercel.com)
- Announced: 2025-08-13T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/how-coxwave-delivers-genai-value-faster-with-vercel
- Record: https://forck.live/items/20838-how-coxwave-delivers-genai-value-faster-with-vercel
- Subject: v0 / AI SDK / AI Gateway

Coxwave reduced deployment times by 85% and increased production deploy frequency from once a week to twice per day by adopting Next.js and Vercel. The company also achieved a 52% decrease in production recovery time with Instant Rollback. Vercel's Preview Environments and AI SDK enabled faster prototyping and release of their AI assistant, Exploration Copilot.

### Vercel — The real serverless compute to database connection problem, solved

- Company: Vercel (vercel.com)
- Announced: 2025-08-13T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/the-real-serverless-compute-to-database-connection-problem-solved
- Record: https://forck.live/items/20837-the-real-serverless-compute-to-database-connection-problem-solved
- Subject: v0 / AI SDK / AI Gateway

Vercel's post explains that the real problem with serverless compute and database connections is not the number of connections during normal operation, but connection leaks caused when functions are suspended. It describes how Vercel Fluid Compute's waitUntil feature can close idle connections before suspension, and provides a one-line code example using attachDatabasePool from @vercel/functions.

### Vercel — The three types of AI bot traffic and how to handle them

- Company: Vercel (vercel.com)
- Announced: 2025-08-13T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/the-three-types-of-ai-bot-traffic-and-how-to-handle-them
- Record: https://forck.live/items/20836-the-three-types-of-ai-bot-traffic-and-how-to-handle-them
- Subject: v0 / AI SDK / AI Gateway

AI bot traffic is growing across the web. We track this in real-time , and the data reveals three types of AI-driven crawlers that often work independently but together create a discovery flywheel that many teams disrupt without realizing it.

### Amp — Diagnostic-Driven Completions

- Company: Amp (ampcode.com)
- Announced: 2025-08-13
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/diagnostic-driven-completions
- Record: https://forck.live/items/7819-diagnostic-driven-completions
- Subject: Amp

Amp Tab now uses language server diagnostics to suggest follow-up edits that resolve errors from previous edits, enabling completions across the entire file. This is particularly useful for changes that introduce fixable errors, such as changing function signatures or renaming variables.

### Amp — 432,000 Tokens

- Company: Amp (ampcode.com)
- Announced: 2025-08-13
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/432k-tokens
- Record: https://forck.live/items/7818-432-000-tokens
- Subject: Amp

Amp increased its context window to 432,000 tokens for Claude Sonnet 4, more than double the previous limit, and rolled it out to all users in the latest CLI and editor extension. The company plans to support 1 million tokens soon.

### Stability AI — Stability AI and NVIDIA Bring Faster Performance and Simplified Enterprise Deployment with the Stable Diffusion 3.5 NIM

- Company: Stability AI (stability.ai)
- Announced: 2025-08-12T19:03:26+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://stability.ai/news-updates/stability-ai-and-nvidia-bring-faster-performance-and-simplified-enterprise-deployment-with-the-stable-diffusion-35-nim
- Record: https://forck.live/items/17484-stability-ai-and-nvidia-bring-faster-performance-and-simplified-enterprise
- Subject: Stable Diffusion

Stability AI and NVIDIA launched the Stable Diffusion 3.5 NIM microservice, a containerized deployment package for SD3.5 Large that delivers 1.8x speed gains over standard PyTorch on H100 GPUs. The NIM bundles the model with Depth and Canny ControlNets in a single container and supports enterprise Ada and Blackwell GPUs. Model weights are available on Hugging Face under the Stability AI Community License, with enterprise licensing for organizations over $1M annual revenue.

### Supabase — Supabase Auth: Build vs. Buy

- Company: Supabase (supabase.com)
- Announced: 2025-08-12T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/supabase-auth-build-vs-buy
- Record: https://forck.live/items/21340-supabase-auth-build-vs-buy
- Subject: Supabase

Authentication appears in nearly every application but is rarely the core value proposition. Yet development teams often spend weeks or months building, testing, and maintaining auth systems.

### Hugging Face — Neural Super Sampling is here!

- Company: Hugging Face (huggingface.co)
- Announced: 2025-08-12T14:52:08+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/Arm/neural-super-sampling
- Record: https://forck.live/items/1637-neural-super-sampling-is-here
- Subject: Platform

Arm released Neural Super Sampling (NSS), an AI-powered upscaling model for mobile graphics, available for developers to experiment with.

### Vercel — Cutting delivery times in half with v0

- Company: Vercel (vercel.com)
- Announced: 2025-08-12T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/cutting-delivery-times-in-half-with-v0
- Record: https://forck.live/items/20839-cutting-delivery-times-in-half-with-v0
- Subject: v0 / AI SDK / AI Gateway

Ready.net, a platform for utility companies, adopted Vercel's v0 to accelerate its design-to-delivery process. The company reports a 50% decrease in time to market and a 30% increase in personal productivity, with feature delivery time cut in half.

### Harvey — Introducing Ask LexisNexis®

- Company: Harvey (harvey.ai)
- Announced: 2025-08-12T12:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/introducing-ask-lexisnexis
- Record: https://forck.live/items/17686-introducing-ask-lexisnexis
- Subject: Harvey

Harvey introduced Ask LexisNexis, a feature that allows users to ask legal questions in natural language and receive AI-generated answers with citations from LexisNexis, grounded in U.S. primary law content including federal and state case law, statutes, and administrative codes. The feature is available in open beta for all Harvey users and includes Shepard's Citations Service to verify the validity of cited authorities.

### Genspark — Introducing Genspark AI Meeting Notes - The World First AI Meeting Notes Tool for Apple Watch Also Available on iOS and Android Apps

- Company: Genspark (genspark.ai)
- Announced: 2025-08-12
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.genspark.ai/blog/genspark-ai-meeting-notes
- Record: https://forck.live/items/18272-introducing-genspark-ai-meeting-notes-the-world-first-ai-meeting-notes-tool
- Subject: Genspark

### LG AI Research — LG AI Research and The Jackson Laboratory Unveil AI-Driven Breakthroughs in Disease Research

- Company: LG AI Research (lgresearch.ai)
- Announced: 2025-08-12
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/news/view?seq=580
- Record: https://forck.live/items/11001-lg-ai-research-and-the-jackson-laboratory-unveil-ai-driven-breakthroughs-in
- Subject: EXAONE

LG AI Research and The Jackson Laboratory are collaborating on AI-driven Alzheimer's disease research, with six joint papers accepted at ICML 2025 and AAIC 2025. The papers cover topics from AI-based behavioral analysis in mice to multi-omics prediction and biological interpretation in humans. The teams are also working on selecting and validating candidate genes related to Alzheimer's.

### Luma — Introducing Modify with Instructions

- Company: Luma (lumalabs.ai)
- Announced: 2025-08-12
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lumalabs.ai/news/modify-with-instructions
- Record: https://forck.live/items/7982-introducing-modify-with-instructions
- Subject: Ray / Dream Machine

Luma introduced Modify with Instructions, a new feature in Dream Machine that allows users to use natural language to direct changes across VFX, advertising, film, and design workflows. The feature is available to all Dream Machine subscribers.

### Amp — Model Evaluation

- Company: Amp (ampcode.com)
- Announced: 2025-08-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ampcode.com/news/model-evaluation
- Record: https://forck.live/items/7820-model-evaluation
- Subject: Amp

The Amp team evaluated GPT-5 over a week and shared their impressions, noting its strengths in coding and instruction following but weaknesses in speed, tool use, and persistence. They also made GPT-5 available for testing in Amp and later used it as the oracle model.

### LG AI Research — LG AI Research and The Jackson Laboratory Unveil AI-Driven Breakthroughs in Disease Research

- Company: LG AI Research (lgresearch.ai)
- Announced: 2025-08-12
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/blog/view?seq=580
- Record: https://forck.live/items/4781-lg-ai-research-and-the-jackson-laboratory-unveil-ai-driven-breakthroughs-in
- Subject: EXAONE

LG AI Research and The Jackson Laboratory are collaborating on AI-driven Alzheimer's disease research, with six joint papers accepted at ICML 2025 and AAIC 2025. The papers cover topics such as cell-type-aware pooling for scRNA-seq, multi-omics integration with MOIRA, and cognitive function prediction in mice. The partnership aims to uncover root causes of Alzheimer's and advance precision medicine.

### Hugging Face — 🇵🇭 FilBench - Can LLMs Understand and Generate Filipino?

- Company: Hugging Face (huggingface.co)
- Announced: 2025-08-12
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/filbench
- Record: https://forck.live/items/1639-filbench-can-llms-understand-and-generate-filipino
- Subject: Platform

Hugging Face announces FilBench, a comprehensive evaluation suite for assessing LLM performance on Tagalog, Filipino, and Cebuano across four categories: Cultural Knowledge, Classical NLP, Reading Comprehension, and Generation. The researchers evaluated more than 20 state-of-the-art models, finding that region-specific LLMs like SEA-LION and SeaLLM are parameter-efficient but still lag behind GPT-4o, and that translation tasks remain challenging for current models.

### Hugging Face — TextQuests: How Good are LLMs at Text-Based Video Games?

- Company: Hugging Face (huggingface.co)
- Announced: 2025-08-12
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/textquests
- Record: https://forck.live/items/1638-textquests-how-good-are-llms-at-text-based-video-games
- Subject: Platform

Hugging Face introduces TextQuests, a benchmark for evaluating LLMs on text-based video games, testing long-context reasoning and learning through exploration.

### OpenAI — Scaling accounting capacity with OpenAI

- Company: OpenAI (openai.com)
- Announced: 2025-08-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/basis
- Record: https://forck.live/items/494-scaling-accounting-capacity-with-openai
- Subject: GPT / ChatGPT / API

Basis built AI agents using OpenAI models including o3, o3-Pro, GPT-4.1, and GPT-5 to help accounting firms save up to 30% of their time and expand capacity for advisory and growth.

### OpenAI — OpenAI’s letter to Governor Newsom on harmonized regulation

- Company: OpenAI (openai.com)
- Announced: 2025-08-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/global-affairs/letter-to-governor-newsom-on-harmonized-regulation
- Record: https://forck.live/items/493-openai-s-letter-to-governor-newsom-on-harmonized-regulation
- Subject: GPT / ChatGPT / API

OpenAI sent a letter to California Governor Gavin Newsom advocating for harmonized state-level AI regulation that aligns with national and emerging global standards.

### Harvey — Harvey’s iManage Integration

- Company: Harvey (harvey.ai)
- Announced: 2025-08-11T23:09:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harveys-imanage-integration
- Record: https://forck.live/items/17687-harvey-s-imanage-integration
- Subject: Harvey

Harvey launched the open beta of a direct integration with iManage, allowing legal teams to securely import documents from iManage into Harvey's Assistant, Workflows, and Vault without third-party middleware. Users can also export documents drafted in Harvey back into specific iManage matters with full versioning, metadata, and audit trails. The integration was tested with Ropes & Gray, Blank Rome, and CMS during early access.

### Vercel — v0.dev -> v0.app

- Company: Vercel (vercel.com)
- Announced: 2025-08-11T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/v0-app
- Record: https://forck.live/items/20840-v0-dev-v0-app
- Subject: v0 / AI SDK / AI Gateway

Vercel rebranded v0.dev to v0.app, introducing an agentic AI builder that can research, reason, debug, and plan, and can collaborate with users or handle work end-to-end. The new version automatically checks work, gives design inspiration, searches the web, plans tasks, reads files, inspects live sites, manages to-do lists, and implements integrations. It is available for free at v0.app.

### Z.ai — GLM-4.5V

- Company: Z.ai (z.ai)
- Announced: 2025-08-11
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://docs.z.ai/release-notes/new-released#2025-08-11
- Record: https://forck.live/items/4460-glm-4-5v
- Subject: GLM

Z.ai launched GLM-4.5V, a 100B-scale open-source vision reasoning model that supports video understanding, visual grounding, and GUI agents, and added a new thinking mode.

### Together AI — OpenAI's New Open gpt-oss Models vs o4-mini: A Real-World Comparison

- Company: Together AI (together.ai)
- Announced: 2025-08-11
- Category: model-update
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/openais-new-open-gpt-oss-models-vs-o4-mini-a-real-world-comparison
- Record: https://forck.live/items/2416-openai-s-new-open-gpt-oss-models-vs-o4-mini-a-real-world-comparison
- Subject: Inference platform

Together AI announces support for OpenAI's newly open-sourced gpt-oss models (20B and 120B parameters) on its platform and publishes a comparison of gpt-oss-120B against o4-mini across five practical tests, claiming gpt-oss-120B wins 4.5/5 vs 3/5.