{"schema":"sanctity-signal/0.2","record_type":"signal","id":"nanbeige-33b-open-weights-e062","domain":"labs","drop_id":"signal-e062","occurred_at":"2026-10-25","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Nanbeige LLM Lab","action_type":"publication of open weights for a 33 billion parameter dense model and its base version under Apache 2.0, with a self-measured comparison table against 3 named rival models","observed":"Sunday October 25, 2026 is the publication date. We read the model card at huggingface.co/Nanbeige/Nanbeige4.3-33B directly; the repository showed an update about 16 hours before our read, and MarkTechPost dates the release to October 25. The publisher is the Nanbeige Team, shown on the page as Nanbeige LLM Lab, with the contact address nanbeige@kanzhun.com; Kanzhun is the company behind the recruitment platform BOSS Zhipin. The licence on the page is apache-2.0. The card gives 33 billion parameters, dense, no mixture-of-experts routing, BF16 weights, and \"a context length of up to 524,288 tokens (512K)\". The configuration is LlamaForCausalLM with 52 layers, hidden size 7,168, 56 query heads and 8 key-value heads, head dimension 128, feed-forward size 22,016, a 166,912-token vocabulary and a RoPE theta of 120,000,000. A base version, Nanbeige4.3-33B-Base, was published with it, and the lab says a 3 billion parameter distillation is planned. The card claims \"state-of-the-art agentic performance among open models in the ~30B parameter class\" and that the model \"leads on most of the benchmarks\", reporting 17 of 20 rows against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at their highest reasoning effort. Selected figures: Terminal-Bench 2.1 at 80.9 against 73.0, 61.8 and 47.2; DeepSWE at 65.5 against 42.5; Terminal-Bench 4.0 at 23.8 against 0.0 for Gemma4-31B; Automation Bench at 54.2 against 38.3; JobBench at 44.8 against 34.5; AIME 2026 at 95.2; YC-bench simulated earnings of $1,025,437 against $641,290. It reports losing 3 rows: SWE-Bench Pro at 60.3 behind 61.7, GDPval at 84.8 behind 86.4, and GPQA Diamond at 89.5 behind 91.3. Reported, not read by us: MarkTechPost notes that Meta publishes Muse-Glimmer-30B at 51.2 on SWE-Bench Pro and 51.7 on Terminal-Bench 2.1, and that Google lists Gemma4-31B at 84.3 on GPQA Diamond, all different from Nanbeige's table. Not published: training data, training compute, post-training method, any independent evaluation, and any hosted interface or price.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Nanbeige4.3-33B model card, Hugging Face, read Oct 26, 2026: licence, architecture and the full benchmark table","url":"https://huggingface.co/Nanbeige/Nanbeige4.3-33B"},{"role":"corrob","label":"MarkTechPost, Oct 25, 2026: Nanbeige releases Nanbeige4.3-33B, with the conflicting competitor figures","url":"https://www.marktechpost.com/2026/10/25/nanbeige-releases-nanbeige4-3-33b-open-agentic-coding-model/"},{"role":"corrob","label":"DataStudios, Oct 2026: Nanbeige4.3-33B, Apache 2.0 dense agentic model with 512K context and benchmarks","url":"https://www.datastudios.org/post/nanbeige-4-3-33b-open-weight-dense-agentic-model-512k-context"},{"role":"corrob","label":"AIWeekly, Oct 2026: Nanbeige opens 33B model under Apache 2.0 with 512K context","url":"https://aiweekly.co/alerts/nanbeige-opens-33b-model-under-apache-20-with-512k-context"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Low","persistence":"Years","rationale":"Cost is medium. Training a 33 billion parameter model and a base version is real spending, undisclosed in amount, and giving the weights away forgoes any licence revenue. Commitment is high: published weights under Apache 2.0 cannot be recalled, and a company that puts its recruitment platform's contact address on a frontier-adjacent release has tied its name to the claim. Reversibility is low for the same reason, because every copy downloaded stays downloaded. Persistence is years: the file outlives the benchmark table, and a permissive licence means derivatives will exist long after anyone checks the scores."},"shift":{"prior":"China's open models come from the big labs and the big platforms.","action":"Sunday October 25, 2026 is the publication date. We read the model card at huggingface.co/Nanbeige/Nanbeige4.3-33B directly; the repository showed an update about 16 hours before our read, and MarkTechPost dates the release to October 25. The publisher is the Nanbeige Team, shown on the page as Nanbeige LLM Lab, with the contact address nanbeige@kanzhun.com; Kanzhun is the company behind the recruitment platform BOSS Zhipin. The licence on the page is apache-2.0. The card gives 33 billion parameters, dense, no mixture-of-experts routing, BF16 weights, and \"a context length of up to 524,288 tokens (512K)\". The configuration is LlamaForCausalLM with 52 layers, hidden size 7,168, 56 query heads and 8 key-value heads, head dimension 128, feed-forward size 22,016, a 166,912-token vocabulary and a RoPE theta of 120,000,000. A base version, Nanbeige4.3-33B-Base, was published with it, and the lab says a 3 billion parameter distillation is planned. The card claims \"state-of-the-art agentic performance among open models in the ~30B parameter class\" and that the model \"leads on most of the benchmarks\", reporting 17 of 20 rows against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at their highest reasoning effort. Selected figures: Terminal-Bench 2.1 at 80.9 against 73.0, 61.8 and 47.2; DeepSWE at 65.5 against 42.5; Terminal-Bench 4.0 at 23.8 against 0.0 for Gemma4-31B; Automation Bench at 54.2 against 38.3; JobBench at 44.8 against 34.5; AIME 2026 at 95.2; YC-bench simulated earnings of $1,025,437 against $641,290. It reports losing 3 rows: SWE-Bench Pro at 60.3 behind 61.7, GDPval at 84.8 behind 86.4, and GPQA Diamond at 89.5 behind 91.3. Reported, not read by us: MarkTechPost notes that Meta publishes Muse-Glimmer-30B at 51.2 on SWE-Bench Pro and 51.7 on Terminal-Bench 2.1, and that Google lists Gemma4-31B at 84.3 on GPQA Diamond, all different from Nanbeige's table. Not published: training data, training compute, post-training method, any independent evaluation, and any hosted interface or price.","update":"A hiring company's lab now ships one, with its own scoreboard."},"inference":{"established":["The weights are on Hugging Face under the Apache 2.0 licence.","The model is dense, 33 billion parameters, with a 512,000 token context.","A base model for custom training was published alongside the instruct model."],"unknown":["That it leads. Every score in the table is measured by the publisher.","What trained it. No data, compute or post-training method is disclosed.","That anyone hosts it. No inference provider served it at publication."]},"narrative":{"hook":"A Chinese recruitment firm's lab published a 33 billion parameter coding model free.","plain":"Nanbeige LLM Lab released Nanbeige4.3-33B on Hugging Face under Apache 2.0, with a 512,000 token context. Caveat: every benchmark number in the table is the lab's own, and some contradict Meta's and Google's published figures.","story":["Open weights are cheap to publish, and the comparison table comes with them. Nanbeige scored its 3 rivals itself, and 3 of those scores do not match what Meta and Google published for their own models. Nobody outside the lab has reproduced any of it."],"trail":[{"d":"OCT 15","t":"Meta publishes 816 billion parameters of model weights free to download, carded in Edition 052. Western open weights and Chinese open weights are now competing on the same download pages."},{"d":"OCT 16","t":"China's Ministry of Commerce and Ministry of Science and Technology propose export controls on 23 AI technologies and open pre-export registration, carded in Edition 053. Technologies already public through open-source release are excluded, which makes publishing weights the compliant route out."},{"d":"OCT 17","t":"DeepSeek releases a 763 billion parameter model free under the MIT licence, with its own search engine, carded in Edition 054. The reference point for Chinese open weights is a frontier-scale model, not a 30 billion parameter one."},{"d":"OCT 23","t":"Moonshot's Kimi K3.1 Pro arrives and is followed by allegations that influencers were paid to promote it without disclosure, recorded in Edition 060's Undercurrents. The credibility of a release claim is now part of the release."},{"d":"OCT 25","t":"Nanbeige LLM Lab publishes Nanbeige4.3-33B and its base version on Hugging Face under Apache 2.0, with a 512,000 token context and a comparison table it measured itself against 3 named rivals."},{"d":"Reacts","r":true,"t":"Qwen, Meta and Google, whose models are scored in the table at numbers they did not publish; Hugging Face, which hosts the claim; BOSS Zhipin's own investors, now funding a frontier-adjacent laboratory; the Ministry of Commerce, whose draft catalogue exempts open-source release; benchmark maintainers asked to arbitrate; and every buyer choosing a 30 billion parameter coding model this quarter."}],"glossary":[{"term":"Open weights","def":"The trained model file itself, published for download. Anyone can run it on their own machines, change it and ship it, without asking the publisher."},{"term":"Apache 2.0","def":"A permissive licence allowing commercial use, modification and redistribution, including inside closed products, with a patent grant and no obligation to publish changes."},{"term":"Dense model","def":"A model that uses all its parameters on every token. A mixture-of-experts model uses a fraction, so it is cheaper to run but larger to store."},{"term":"Agentic benchmark","def":"A test where the model must use tools and take many steps, such as fixing a repository or working a terminal, rather than answering one question."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Open weights are China's strategy","text":"A second tier of Chinese firms now publishes frontier-style models for free.","abstain":false},{"id":"b","label":"A recruiting company's advertisement","text":"A hiring platform gets attention and engineers from a free model release.","abstain":false},{"id":"c","label":"Benchmarks are the marketing","text":"Self-measured tables that beat rivals are the cheapest way to be noticed.","abstain":false},{"id":"d","label":"The normal open model cadence","text":"Small laboratories ship 30 billion parameter models monthly. This one is routine.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"c","direction":"weakens","observable_event":"An independent evaluation, by a benchmark maintainer or a third-party harness, reproduces Nanbeige's Terminal-Bench 2.1 and DeepSWE figures within 3 points, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"Nanbeige revises or withdraws the comparison table, or Meta or Google publicly disputes the figures attributed to Muse-Glimmer-30B or Gemma4-31B, by January 31, 2027.","window_end":"2027-01-31","state":"unresolved"},{"linked_interpretation":"a","direction":"strengthens","observable_event":"At least 3 more Chinese companies outside the known AI laboratories publish open weights above 20 billion parameters under permissive licences, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"}]}