Open Weights
10 items across 4 editions · appeared in the last 2 editions in a row. First seen Fri 11 Sep, last seen Tue 15 Sep.
- The paper, published in Nature Medicine on 15 September 2026, reports that a fully on-premise clinical agent "achieved 90.04% accuracy on a seven-disease task and 83.8% accuracy on a four-disease task" across two MIMIC-IV-derived benchmarks. On the primary benchmark MIRA-v2, Qwen-3.5 reached 90.0%, GLM-5 89.7%, GLM-4.5-Air 88.4% and GPT-OSS 85.3%, against a cloud baseline of GPT-5.2 at 90.7% — "The best on-premise model was, therefore, within 0.7 percentage points of the cloud baseline."
- On the CDM benchmark (four abdominal categories, n = 2,400) Qwen-3.5 scored 83.8% and GLM-4.5-Air 81.2%; the paper states "The highest previously reported open-weight result on this benchmark was 70.5% (Gemma-3)."
- The authors report that behavioural consistency across repeated runs discriminated correct from incorrect diagnoses better than the model's own probability score (AUC = 0.860 versus 0.747), and that at a consistency threshold of 0.90, "49.4% of cases were retained at 98.9% diagnostic accuracy" — 272 cases routed to autonomous handling, with three errors among them.
- The paper states its own limits plainly: both primary benchmarks derive from MIMIC-IV and "a single-institution data ecology"; the evaluation is text-only; consistency thresholds "must be calibrated to the deployment configuration"; and "all evaluations were retrospective simulations", with prospective studies and bias audits still required.
- "Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models" (arXiv 2609.12586, submitted 11 September, announced 14 September) is by Daan van der Weijden, Nathan Brack and Selene Báez Santamaría; the HTML version lists University of Zurich addresses.
- The paper reproduces the original subliminal-learning experiments — in which a teacher model transmits behavioural preferences through semantically unrelated data — across two trait types, animal preferences and misalignment, and three modalities: number sequences, code and chain of thought. It then extends them with new preference categories, a chess move-generation task and the open-weight model Ministral8B.
- The abstract states: "Our reproduction supports the original paper's claims, but our extensions show they are not universal as transmission strength varies across traits and tasks, and one model shows almost no effect at all."
- Preprint, not peer reviewed. The authors say they used open-weight models because "the original paper's GPT-4.x fine-tuning is no longer available", so the reproduction does not re-test the closed models the original result was reported on.
- In a statement released on Sunday 13 September by China's Ministry of Foreign Affairs and reported by CNBC, President Xi Jinping said at the BRICS summit in New Delhi that China "will pioneer the establishment of a BRICS AI open-source community, support the cooperation in developing and applying large language models, hold AI seminars and training courses, and build an open AI ecosystem".
- CNBC reports Xi also said China will work to establish a BRICS digital ecosystem cloud platform and conduct digital skills training, technological exchange and industrial alignment, and proposed a BRICS engineer cultivation alliance and a youth exchange programme for scientific and technological innovation. In separate reporting CNBC says Xi called for a "consensus-based global AI governance framework".
- BRICS was established in 2009 and now comprises 11 nations including China, India, Russia, Iran and the United Arab Emirates, CNBC notes.
- CNBC says Xi did not address the ongoing AI safety debate. No funding figure, timetable or governing structure was announced for the open-source community or the cloud platform, so this is a commitment rather than a launch.
- TechNode Global, reporting a filing made to the Hong Kong stock exchange on 13 September, says Z.ai plans a share placement of HK$15.68 billion (about $1.98 billion) at HK$714 per share — a 9.96% discount to the 11 September close of HK$793 — representing about 4.5% of enlarged share capital and expected to close on 16 September.
- Alongside it the company plans convertible bonds with gross proceeds of $3.016 billion and principal of RMB20.14 billion, zero coupon, maturing in 2027, with an initial conversion price of HK$892.50 per share.
- Of the proceeds, 60% is earmarked for next-generation GLM models and training, inference and computing infrastructure, 15% for business expansion, strategic investments and potential acquisitions, and 25% for capital structure, working capital and general corporate purposes, with deployment expected by 30 June 2028.
- Both transactions remain conditional and had not closed; TechNode notes the placement may not proceed if its conditions are not met. Only this outlet's account of the filing was read for this item.
- The Globe and Mail, citing four sources, reports Cohere is in advanced talks to raise between US$2-billion and US$3-billion at a US$20-billion valuation, including financing from the Canadian government and existing backers, with the German government also in talks to participate.
- That compares with Cohere's roughly US$7-billion valuation in September 2025. The paper notes Mistral's €3-billion raise at a €21-billion valuation, OpenAI at US$852-billion in March 2026 and Anthropic at US$965-billion in May 2026.
- The round would be the largest on record by a private Canadian startup. The Globe and Mail says it could close as early as next week but that timing could slip.
- Cohere has not confirmed the round and the reporting rests on unnamed sources; no term sheet or filing has been made public.
- The paper, posted to arXiv on 9 September 2026 and announced in the 11 September listing, reports that the system "scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold".
- The abstract states the pipeline "operates entirely in natural language, with no formal prover, external tools, or internet access", using three Nemotron 3 Ultra checkpoints — the general-availability model and two post-trained specialists — in "an iterative search that generates, verifies, and refines candidate proofs", with a separate high-compute stage selecting each submission.
- The authors say they release both post-trained checkpoints plus the training data, training and inference code, the submitted solutions, and "Nemotron-IMO-Bench, a new benchmark of 200 novel olympiad-level problems".
- The arXiv abstract page carries no affiliation block; Hugging Face lists the institution as NVIDIA. The IMO score is the authors' own report of their own submission and is not peer reviewed.
- The paper, posted to arXiv on 10 September 2026, reports that "MoEs degrade more rapidly under data repetition", with the effect growing as sparsity increases: dense 80M-parameter models tolerate "8x" repetition with minimal decline while MoEs "begin to suffer at 4x" and underperform dense alternatives at "32x".
- The study spans models from "80M to 1B active (8.5B total) parameters". The authors are Atindra Jha, Margaret Li, Jure Leskovec, Percy Liang and Luke Zettlemoyer.
- With strong masking-based regularisation, MoEs keep their advantage over dense models "even when data is repeated more than 64 times", though the paper says no method fully recovers all-unique-data performance — a direct constraint on sparse architectures as high-quality text runs short.
- Preprint, not peer reviewed. The largest configuration is 8.5B total parameters, well below frontier scale, and the paper does not claim the thresholds transfer.
- TechCrunch reported on 11 September at 12:35 pm PDT that Moonshot AI is "targeting $2 billion in annualized revenue by the end of the year", a doubling of its August run rate. For scale, TechCrunch puts OpenAI's revenue run rate at $40 billion and Anthropic's annualised revenue at $65 billion.
- OpenRouter data cited by TechCrunch shows Moonshot's K3 models generating "as many as 300 billion tokens being generated each day" on that platform, with usage down slightly in recent months.
- The figures matter because Moonshot ships open weights, which carry lower margins than closed models; a $2 billion run rate would be the strongest commercial evidence yet for that business model.
- In the same week Anthropic accused Moonshot of routing "nearly 300,000 requests from Kimi directly to Claude Opus" and collecting "more than 23 million responses". The revenue figures are Moonshot's own, given to investors, and are not independently verified; Moonshot's response to the distillation allegation is not in the piece.
- "Big Enough to Break Out" (arXiv 2609.10780, submitted 9 September) compares two PentestGPT-based systems: a legacy human-in-the-loop system on open-weight Kimi K2.5, and a newer autonomous system on Claude Opus 4.8. Across three public targets the autonomous system solves all three, including the two the legacy system never finishes.
- The authors flag the legacy result as the more surprising one: even on machines it fails to solve, it completes about half the subtasks while running on ordinary university GPUs with no provider guardrails — a capability floor available to anyone with open weights and campus hardware.
- The paper explicitly declines to attribute the gain, since model, harness, autonomy and memory architecture all changed together. Adding a coverage-memory layer to both systems improved neither, and in reviewable stalled runs the limiting factor looked like planning and commitment rather than lost memory. Three targets is a very small sample.
- Published 10 September, the evaluation measures intelligence targeting and conventional weapons capability across Claude Mythos Preview, Mythos 5, Opus 5 and Sonnet 5, plus open-weights Kimi K3 and GLM 5.2. On 6,000 YFCC100M Flickr images, Mythos Preview reached a 37.0 km median error with 23.7% of images placed within 1 km, against 181 km for Opus 5, 384 km for Sonnet 5 and 385 km for Kimi K3; Anthropic compares this to 151 km for top GeoGuessr players.
- On text geolocation from anonymised GeoText tweets covering 1,697 users, median error ranged from 20.1 km (Mythos Preview) to 31.3 km (Sonnet 5), and 135 users — 8% of the corpus — were reliably placed within 1 km by at least one model. On account linkage across synthetic social media, Mythos Preview processed median 37,000-word samples in about 11 minutes, against roughly 2.5 hours for human analysts.
- On simulated drone terminal guidance against a parked high-visibility vehicle, Opus 5 struck the target on 80% of runs, Mythos Preview 70%, Mythos 5 53%, Kimi K3 15% and Sonnet 5 5%. Across all nine difficulty settings Opus 5 hit on 20% of 540 launches. Under GPS denial, only Opus 5 kept about a third of flights inside five metres.
- Anthropic frames these as capability ceilings for isolated models and notes human teams with internet access would likely do better. The drone work is in simulation, not flight, and the report does not disclose what mitigations follow. The open-weights results matter most: Kimi K3 trails the frontier but is not far behind on photo geolocation, and cannot be withdrawn.