OpenAI’s Open-Weight Model: Release Date, Specs, and the Truth About a Successor

Key Takeaways

• The openai open-weight model story, verified as of October 2026: OpenAI’s current and only open-weight releases are gpt-oss-120b and gpt-oss-20b, launched August 5, 2025 under Apache 2.0. There is no newer release.
• Rumors of a “gpt-oss-2” are pure speculation — no model card, no weights repository, and no OpenAI statement exists. Third-party quantized builds (e.g., Modelers.cn’s October 2026 repacks) are community repackagings, not new models.
• The two models are mixture-of-experts reasoning models: the 120b runs on a single 80GB GPU, the 20b fits in 16GB of memory — consumer-grade hardware.
• OpenAI’s own DevDay 2026 (September 29) contained no open-weights announcement whatsoever, while competitors — Qwen, GLM, Kimi, DeepSeek — have kept shipping open models.
• Sam Altman’s stated position has evolved from a January 2025 admission that OpenAI was “on the wrong side of history” on open source to a July 2026 declaration: “I want the US to win in AI both in open source and proprietary models.”

Introduction

Search for “OpenAI open model release date” right now and you’ll find a fog of speculation: rumored successors, invented version numbers, and breathless posts about releases that never happened. Let’s clear the fog with verified facts.

The honest headline is this: OpenAI has not released a new open-weight model in October 2026, and none has been announced. Its latest open-weight release remains gpt-oss-120b and gpt-oss-20b, launched August 5, 2025 — more than a year ago. An October 2026 analysis put it bluntly: “More Than a Year After gpt-oss, OpenAI Has No Successor. Build As If None Is Coming.”

That absence is itself the story. Here’s what the models are, where to get them, what the rumors actually amount to, and what OpenAI has said about its open-model future.

OpenAI Open-Weight Model Release Date: The Confirmed Facts

Item	Status	Date
gpt-oss-120b and gpt-oss-20b released (Apache 2.0)	Confirmed	August 5, 2025
OpenAI announces intent to release its first open-weights model since GPT-2	Confirmed	~April 2025
Successor / “gpt-oss-2”	Rumored only	No date exists

The August 2025 launch ended a six-year drought: gpt-oss was OpenAI’s first open-weight release since GPT-2 in 2019. It arrived months after Sam Altman’s January 2025 Reddit AMA admission — made in the wake of DeepSeek R1 — that “I personally think we have been on the wrong side of history here and need to figure out a different open source strategy.” In April 2025, the company confirmed it would ship open weights “in the coming months.” It did.

What gpt-oss-120b and gpt-oss-20b Actually Are

The two models are text-only mixture-of-experts (MoE) reasoning models with adjustable reasoning effort and full chain-of-thought access. The naming needs a small decoder ring: they’re branded 120b and 20b, but the actual parameter counts per the Hugging Face model cards are 117B total / 5.1B active (120b) and 21B total / 3.6B active (20b).

The specs that matter for builders:

• Hardware: the 120b model fits on a single 80GB GPU (e.g., an NVIDIA H100); the 20b runs within 16GB of memory — reachable on consumer cards like the RTX 3090, 4090, or 5080.
• Quantization: natively MXFP4-quantized 4-bit weights.
• Context: 128K tokens, via RoPE with alternating full-context and sliding-window attention layers.
• Tokenizer: the same one used by GPT-4o and OpenAI’s other API models.
• Performance (OpenAI’s launch claims): the 120b was pitched as approaching o4-mini-level performance, the 20b comparable to o3-mini. Launch coverage cited Codeforces scores of 2622 (120b) and 2516 (20b) — treat these as OpenAI’s own launch-day numbers, not independently replicated benchmarks.

Where to download and run them

The models are widely available: Hugging Face, Ollama (`ollama run gpt-oss:20b`), vLLM, llama.cpp, LM Studio, Microsoft Azure, AWS Bedrock and SageMaker (notably the first OpenAI models on Bedrock), Fireworks, and Together AI — with day-one optimized support in NVIDIA’s stack.

The license: why Apache 2.0 matters

gpt-oss shipped under Apache 2.0 with a minimal complementary usage policy — a genuinely permissive license with no user-count tiers or gating. That choice is significant: it means startups and enterprises can build on the weights commercially without the friction that encumbers some “open-ish” releases. As OpenAI’s Greg Brockman put it at launch: “One of the things unique about open models is that people can run them locally.” Local, private, offline inference — the thing the API can never offer — is the whole point.

The Successor Rumors: What We Checked

Now the part everyone is actually searching for. Here’s the verified state of every “new open model” claim circulating in October 2026:

• DevDay 2026 (September 29, 2026): No open-weights announcement. The event covered agentic products — Dots agents, GPT-6.1 Sol, Codex Cloud, the Agents API, ChatGPT Spaces — and our own DevDay 2026 coverage confirms zero open-model news. (That post and this one are fully distinct — no overlap.)
• “gpt-oss-2” rumors: Speculation only. The Register mused in July 2026 that a successor might appear the following month — it didn’t. As the October 2026 Glad Labs analysis states: “Until there’s a model card or a weights repository, there’s no release.” Version numbers in this space, the author notes, “get invented by blog posts constantly.”
• Modelers.cn builds (October 2026): New W8A8 and 4-bit quantized builds of gpt-oss-20b for Huawei’s Ascend ecosystem appeared around October 14, 2026. These are third-party community repackagings of OpenAI’s existing weights — useful for Ascend hardware owners, but not new models. Do not mistake a quantization for a release.

Meanwhile, the competitive landscape has not stood still. While OpenAI has shipped nothing new in open weights for 14+ months, Qwen, GLM, Kimi, and DeepSeek have kept releasing open-weight models — a contrast worth sitting with.

What OpenAI Has Actually Said

Stripping away rumor, OpenAI’s on-the-record open-model strategy looks like this:

• January 2025 (Reddit AMA): “I personally think we have been on the wrong side of history here and need to figure out a different open source strategy.” — Sam Altman
• April 2025: Formal announcement of forthcoming open weights “in the coming months.”
• ~July 2026: Altman backed the industry letter to the U.S. administration calling for a thriving open-weight AI ecosystem, posting on X: “I want the US to win in AI both in open source and proprietary models.”
• August 2026 (Politico interview): Altman argued society should accept “some bad things” — hacks, scams — because the good from open models will outweigh them by orders of magnitude, warning against over-restriction concentrating power in too few hands.

The through-line: OpenAI’s leadership now rhetorically champions open weights while shipping exactly one open-weight release in over six years — and nothing since. Rhetoric and release cadence are telling different stories.

Why “No Successor” Is Itself a Story

There are a few plausible readings of the silence, and it’s worth distinguishing what we know from what we’re inferring. What we know: no roadmap, announcement, or commitment exists. What analysts infer: frontier training runs are enormously expensive, and each open release potentially cannibalizes API revenue while arming competitors — the economic incentives for a lab of OpenAI’s structure point toward proprietary releases, whatever the rhetoric.

For builders, the practical conclusion from the October 2026 analysis is blunt and useful: build as if no successor is coming. The gpt-oss models remain capable, permissively licensed, and runnable on local hardware today. Betting a product roadmap on a rumored OpenAI open release is betting on something with no model card, no repository, and no date.

What This Means for You

• If you want open weights from OpenAI today: gpt-oss-120b and gpt-oss-20b are real, available, and Apache 2.0-licensed. Start with the 20b on Ollama if you have a decent GPU.
• If you’re waiting for gpt-oss-2: stop waiting and evaluate what’s actually shipping — from OpenAI’s competitors as well as from OpenAI itself.
• If you’re choosing a strategy: weigh the trade-off explicitly. Open weights give you local control, privacy, and no API bills; proprietary frontier models give you the best raw capability. The right answer depends on your constraints, not on hype.

For related reading, see our OpenAI DevDay 2026 announcements breakdown, the Abacus AI Smaug open-models overview, and our analysis of real 2026 costs for OpenAI and Anthropic’s cheaper models.

The open-weight chapter of OpenAI’s story is real but short: one release, August 2025, still the latest. Everything else you’ve read about a successor is, as of today, a rumor wearing a headline’s clothes.