2026-08-30

1. The largest open-weight model this wiki holds shipped with the fewest strings, from a lab that had no page here — score 2.00

• Tencent Hunyuan released Hy4 preview on 2026-08-28 and open-sourced it the same day — 770B total parameters, 49B activated, MoE over 78 layers (77 sparse, 256 routed experts + 1 shared, top-8 routing), one native MTP layer (10B/0.7B) for speculative decoding, context over 1M tokens. Apache 2.0 on both the BF16 checkpoint and an FP8 quantisation, mirrored to ModelScope, GitCode and CNB; API at $0.834/M input · $2.501/M output through Tencent Cloud TokenHub and OpenRouter (source)
• Vendor figures: Terminal Bench 2.1 85.4, DeepSWE 28.0 → 64.3, and an internal blind evaluation of 203 engineering tasks scored by 163 Tencent experts placing it at 2.99/4 against Kimi K3's 2.94 and GLM-5.3's 2.92 — 46.8% wins / 12.8% ties / 40.4% losses head-to-head against GLM-5.3
• Why it matters: three Chinese open-weight releases in fifteen days, and the licences separate more cleanly than the models do — MIT for GLM-5.3-Flash, qwen-community-1.0 for Qwen3.8-Flash-Next, a $10 billion-revenue security-review clause for GLM-5.3, and Apache 2.0 flat for the biggest artefact of the four. The largest model carries the loosest terms
• What is not established, and it is a lot: no harness is named for any benchmark row; the blind evaluation was designed, run and scored by the party being measured, and its margin is 0.05 and 0.07 on a 4-point scale with no variance or inter-rater figure published; and no third party has measured this model at all — it is absent from today's Artificial Analysis capture, which lists Tencent only through the earlier Hy3 at Index 42. Nothing here was read first-party: tencent.com and huggingface.co are both blocked from this sandbox
• One checkable claim does not check out. The 85.4 is reported as surpassing DeepSeek V4-Pro; this wiki has carried a vendor-stated 87.9 for DeepSeek V4-Pro-0813 since 2026-08-14, also with no published harness. Recorded as a conflict on both sources, not resolved
• → Hy4 preview (new), Tencent (new), Open-Weights Policy Fight

2. A leaderboard changed what it measures, and every model's headline number moved without anyone touching a model — score 1.56

• Hugging Face added Indian-English to the Open ASR Leaderboard as Voice Arena Monsoon, placed in the default column set rather than behind an opt-in toggle, so it now contributes to the headline Average WER of every listed model. Public and private Hindi join the Multilingual tab, where a model ranks only if it supports every selected language. Described as the leaderboard's first Global South language (source)
• Why it matters: the default placement is the whole substance. An opt-in column changes nothing for a model that does not opt in; a default column re-scores the entire board with nothing retrained and nothing resubmitted. It is the Eval Harness Configuration problem moved up a level — from which harness ran this to which languages count as the average
• What is missing: no before/after table, no model whose rank moved, no figure for how far Average WER shifted. The re-scoring happened and its size is unpublished
• → Hugging Face, Eval Harness Configuration

3. Maintainers say the exploit now arrives before the patch, and the evidence for it is anecdote — score 1.50

• Anil Madhavapeddy — Cambridge computer scientist and core maintainer of the OCaml compiler — reports exploits attempted within minutes of a patch being shared for discussion, with the input being the rumour of a bug rather than the diff. rclone's maintainer reports going from about 20 security disclosures in ten years to over 40 in the last month. The summary claim: mean time to exploit is now −7 days. Surfaced via Simon Willison, 2026-08-28 (source)
• Why it matters: every other item on AI-Enabled Cyberattacks measures offensive capability inside a controlled evaluation — CyberGym, ExploitGym, Chrome V8 counts. This is the first one reporting the effect on ordinary open-source maintenance, which is where that capability lands if it lands anywhere
• Read the weakness with the claim: no methodology is published for −7 days — no sample, no definition, no window; no model, agent or harness is named, so the attribution to coding agents is the authors' reading of a pattern rather than a measurement; and disclosures are not exploits. It is carried as a signal with its limits stated, not as a finding
• → AI-Enabled Cyberattacks, Agents (LLM Agents)

4. A sentence this wiki published four days ago became false, and a committed leaderboard is what caught it — score 1.20

• Qwen3.8-Flash-Next read "no third party has measured this model at all". Artificial Analysis now lists it at Intelligence Index 56, context 256k, Cost per Task $0.10 — absent from the 2026-08-23 capture, present in both 08-29 and 08-30 (source)
• Why it matters: the correction came from a file the pipeline had already committed and nobody had re-read against the page it contradicts. That is the cheapest class of error this system can find and the easiest to leave standing
• It settles less than it appears to. Artificial Analysis publishes a composite, not SWE-bench Pro or JobBench — so the 19.1-point JobBench margin that page flagged for independent reading is still unread by anyone outside Alibaba. At 56 the model sits one point under GLM-5.3-Flash (57) on the only scale both now appear on
• → Qwen3.8-Flash-Next, GLM-5.3-Flash

5. OpenAI cut off the biggest AI coding product it does not own, and the reason is contractual rather than safety — score 1.09

• Our decision on Cursor following its acquisition by SpaceX (2026-08-28) ends Cursor's direct access to OpenAI models on 2026-11-12. OpenAI states it cannot be confident SpaceX will use the technology within its terms of service, citing the Twitter acquisition and xAI having admitted violating those terms (source)
• Cursor CEO Michael Truell puts OpenAI's share at about 5% of Cursor's traffic; Anthropic is reported to have answered with promises of more Claude support
• Why it matters: the gates this wiki tracks on GPT-5.6-Cyber and Astra restrict what a model may do. This one restricts who may resell it, and it is the first instance held here of a frontier lab revoking distribution for contractual reasons. The xAI side is the less obvious half — three Grok models on that page are recorded as Cursor-trained, so the acquisition is the source of the workflow data xAI's coding line was built on, and owning it now costs the product a supplier
• On the score, because it is the day's biggest external story and sits last. Scored official_blog 1.2 × business deals 0.7 × OpenAI 1.3. Read instead as an agents/tool-use item — Cursor is agent tooling, interests.md weights that lane 1.5 — it would compute to 2.34 and lead the brief. That reading was declined: the item is a contract decision about resale rights, not a change in what an agent can do, and interests.md names "changes to agents/MCP standards" as the tracked signal in that lane. The alternative is printed so the choice can be argued with
• → OpenAI, xAI, Model Routing