K2 Horizon
Vintage: 2026-09. Primary evidence is one official @IFM_AI X post dated 2026-09-03, as hydrated in 2026-09-03-x-ai-pass-3-sep-2026-open-model-fleets-agentic-coding-price (
fetch_method: x-mcp). Launch-page / model-card claims (ifm.ai/k2,ifm.ai/blog/k2) are pointers named inside the post, not pages this pass fetched. A same-morning video teaser (Meet K2 Horizon) was deliberately not cited (no transcript).
K2 Horizon
One-line summary: Institute of Foundation Models' 3 Sep 2026 six-model fleet (0.9B–375B); the launch claim is fully open code, training data, and recipes — not a named-frontier benchmark win.
What it is
A "connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters," announced by the Institute of Foundation Models on official @IFM_AI X. Treat this as an issuer X claim. The post names no license, no benchmark table, and no comparison against a named frontier model.
Why it matters to this thread
Open-weight releases and what "fully open" actually grants are in-scope (chinese-open-weight-frontier-parity, open-vs-closed-source-model-economics). This is the first dated IFM / K2 Horizon page in the thread. It does not rewrite the June GLM 5.2 / July Kimi K3 podcast evidence on those pages.
Key facts (from 2026-09-03-x-ai-pass-3-sep-2026-open-model-fleets-agentic-coding-price)
The bullets are one X post (fetch_method: x-mcp). Permalink: https://x.com/IFM_AI/status/2095497035806113861.
- Fleet: six models, 0.9B–375B.
- Per-size-class claim: "top-tier performance in every size class" on coding and agentic tasks; the 0.9B, 3.7B and 7B models "setting new state of the art at their respective scales." The strongest specific claim is at the small end, not at 375B.
- Openness claim (the rhetorical lead): "the largest fully open-source model launch in AI history," with "fully open code, training data and recipes."
- Named artifacts (pointers, not fetched):
ifm.ai/k2,ifm.ai/blog/k2, Hugging Face collection.
What this source does not establish
- No license grant. "Fully open code, training data and recipes" is not a license. Do not characterize K2 Horizon against other open-weight releases until the launch page or model cards are fetched.
- No benchmark table and no named-frontier comparison in the primary post.
- Not a re-file of Kimi K3 or GLM 5.2. Those stay on chinese-open-weight-frontier-parity / moonshot-ai.
- Video teaser not cited. Same-morning @IFM_AI "Meet K2 Horizon" video is excluded (no transcript). Do not add that permalink here.
- @karpathy / @OpenAI / @AnthropicAI had zero original posts in the pulled window — silence, not a fetch miss.
Open questions
- What license actually ships with the code, data, and recipes?
- Do the 0.9B / 3.7B / 7B SOTA-at-scale claims survive a named-benchmark table?
- How does a six-model fleet release compare, as an openness event, to single-weight drops (Kimi K3, GLM 5.2)?