Research Vault

Open-weight LLM Weekly Recap — week of 2026-06-22–2026-06-28

Properties1
tagsresearch, llm, open-weight, weekly-recap

2 min read

Open-weight LLM Weekly Recap — week of 2026-06-22–2026-06-28

🗓 Open-weight LLM week of 2026-06-22–2026-06-28 (1 release)

🏆 Headline: Alibaba’s Qwen team shipped Qwen-AgentWorld-35B-A3B — a 35B-parameter (3B active) MoE that is being framed as the first “language world model” unifying seven agent domains (MCP, Search, Terminal, SWE, Android, Web, OS) in a single training stack (CPT→SFT→RL, 262K ctx). On the AgentWorldBench Overall it posts 56.39, beating Qwen3.5-35B-A3B base (47.73), DeepSeek-V4-Pro (52.97), Kimi K2.6 (53.42), and GLM-5.1 (51.31). The bigger 397B-A17B sister variant climbs to 58.71 — the first open-weight model I’ve seen cleanly clear GPT-5.4 (58.25) and Claude Opus 4.8 (56.59) on an agentic benchmark.

📊 Releases (ranked by impact)

─────────────── Alibaba Qwen-AgentWorld-35B-A3B (35B params, 3B active MoE, Apache 2.0, 2026-06-23)

  • Key: AgentWorldBench Overall 56.39 · AgentWorldBench@397B-A17B 58.71 · 262K ctx
  • Beats: Qwen3.5-35B-A3B base (47.73), DeepSeek-V4-Pro (52.97), Kimi K2.6 (53.42), GLM-5.1 (51.31) on AgentWorldBench Overall; the 397B sister variant beats GPT-5.4 (58.25) and Claude Opus 4.8 (56.59)
  • Loses to: n/a within the open-weight cohort — leads the field
  • Why it matters: First open-weight release to frame “language world model” as a unifying pretraining objective across 7 agent domains, with the larger variant becoming the first open-weight model to top GPT-5.4 and Claude Opus 4.8 on a published agentic benchmark
  • 🔗 HF: Qwen/Qwen-AgentWorld-35B-A3B · Blog · arXiv 2606.24597

📈 Trends this week

  • MoE continues to dominate the open-weight frontier: the week’s only release is a 3B-active MoE, and the leaderboard slots it touches (Qwen3.5-35B-A3B, DeepSeek-V4-Pro, Kimi K2.6, GLM-5.1) are all MoE — five-for-five on the top of AgentWorldBench.
  • Apache 2.0 wins again: Qwen’s choice keeps the Qwen family at the most permissive end of the open-weight spectrum, ahead of Llama Community-style restrictions.
  • Chinese labs collectively own the agentic-benchmark leaderboard: Qwen (35B + 397B), DeepSeek-V4-Pro, Kimi K2.6, GLM-5.1 — all from PRC orgs, all beating Western closed-API comparators on at least one agentic axis.

🔭 Looking ahead

The 397B-A17B sister variant is the one to watch — if Alibaba releases it under Apache 2.0 on the same schedule, this becomes the first genuinely open-weight model that beats GPT-5.4 and Claude Opus 4.8 on a public agentic benchmark. Worth re-checking the Qwen HF org next week for a separate model card and a v2 AgentWorldBench leaderboard.

Tracker: 1 total cumulative release in file · file at /home/pi/projects/llm-tracker/releases.md

See also

Graph View