This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

French Mistral put all weights of its flagship Mistral Large 4 (internal codename "Le Chonk") on Hugging Face. This is the open-weight camp's first touch of the trillion-parameter scale: ~1T total (1.05T), ~52B activated per token (4 of 128 experts + 1 shared), 1M context, text plus image input. Trained from scratch in Mistral's own European data center on ~3,800 NVIDIA Grace Blackwells, with three more weeks of RL than the Oct 6 preview - the released checkpoint is a stronger, further-trained version, not the preview.
It's a modified MIT (Mistral Ecosystem Open-Weight License): self-use, fine-tuning, commercial all free; but wrapping it as "model-as-a-service" to sell API requires a separate commercial license from Mistral. That means fewer third-party hosts and no price war - to use it cheaply, either go through Mistral's own API / Microsoft Foundry / OpenRouter, or self-host (community GGUF quants exist; the smallest 2-bit runs on a 512GB unified-memory Mac at ~90 tokens/s prefill, ~15 generation).
Take: this is Europe's first trillion-scale open-weight model, and the play is clear - not fighting the US-China "overall intelligence" battle, but targeting "sovereign deployment + coding savings," the pocket enterprises actually pay for. Our plaza has listed mistral-large-4, 6 sellers, deepest discount ~90% (an API snapshot at publish; actual price follows the plaza in real time).
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| GO-3 Preview + WITA-OMNI 1.0 + Lingqu OS 1.0 (Zhiyuan) | 10-29 | Embodied-AI suite (open by year-end) | The new embodied base model GO-3 Preview uses the native WRAM architecture, the world's first embodied model with full-chain scaled learning across pre-training, post-training, and post-deployment; first on RoboDojo / RoboTwin / RoboColiseum, fully open by year-end. WITA-OMNI fuses vision, audio, language, expression, and body motion across 100+ languages and 29 dialects. Lingqu OS 1.0, the industry's first embodied-native OS, ships alongside and is donated to the OpenAtom Foundation. Also proposes the "Worken" metric (robots x autonomous hours x output efficiency). All vendor-reported; no independent replication. |
| Wenxin Unlimited OCR 2 (Baidu, open source) | 10-30 | Long-document OCR | Pushes the OCR ceiling from "dozens of pages" to unlimited pages, SOTA on MPDocBench; reports claim 3,600+ page books parsed in one pass with accuracy and throughput. Contracts, case files, and ancient texts can skip the segmentation-splicing work. |
| Nemotron 3.5 Transcribe (NVIDIA, open source) | 10-28 | Speech transcription | ~34B total / ~4.3B active offline voice model: one model covers 40-language transcription, 31-language translation, speaker diarization, and audio Q&A, with 800K hours of data and fine-tuning recipes; weights on Hugging Face. |
| FluxVLA Engine V2 (Zhuji Dynamics, open source) | 10-30 | Embodied VLA train/deploy engine | Spans data, training, eval, and real-robot deployment; V2 adds online RL, 2.3-9.6x faster inference; two SmolVLA task success rates up 10.0 and 15.4 points (vendor). |
| Idol (Speridlabs, open source) | 10-30 | Image-to-3D | Single image to 3D Gaussian splats: 600M params, runs in under 7.2GB VRAM on consumer cards, 1M splats in 3 seconds; weights, SDK, and online demo all Apache 2.0. |
| Apex-35B-V1.0 (Super Intelligence) | 10-30 | General 35B model | Vendor-claimed 20 of 21 benchmarks over Qwen3.6-35B-A3B, 18-1-0 over 40B-class Flint, covering code, agent, deep search, and research, via "recursive self-improvement." All self-reported, unreplicated. |
| Strands Decider 2B + Strands Box (AWS, open source) | 10-30 | Small decision model + sandbox | A 2B lightweight decision model for "pick one from predefined options with confidence," for model routing, tool selection, and guardrails, with a companion sandbox. Reports note concurrency and cold-start remain weak. |
| Droid 1.0 (Factory) | 10-30 | Agent coding engine | Defaults to "programmatic tool calling": an agent makes hundreds of tool calls in one model request, saving tokens and latency; vendor-claimed leading pass rate and cost efficiency on Terminal-Bench 4. |
Today's keyword is "weights + ledgers." Mistral puts a trillion parameters on Hugging Face, OpenAI releases 722 manuscripts and doubled subscription quotas, while the evaluation body holds the "private test set" knife to leaderboard-gamers' throats - more openness, yet trustworthy measurement gets more expensive. For users: stop reading only the vendor card's score; check for independent replication first, then compute each task's real token bill, and only then look at rankings.
(Data from official announcements and public reports, 2026-10-31; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa