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.

Moonshot dropped the first K3.1 model in the early hours: Kimi K3.1 Pro, with a Highspeed variant alongside. One-line summary: parameters didn't grow, the brain got better, and the price dropped. It's trained on the K3 base with a much larger post-training compute budget (joint RL across hundreds of domains), which Moonshot calls "the strongest 3T-class open-weight model yet."
The 14 score charts include real weaknesses: Terminal-Bench Science is only 38.1, far behind GPT-6 Astra's 68.1 - if your agent runs scientific computing in a terminal, this isn't for you yet; its own coding benchmark 61.2 still trails Astra's 66.5; OSWorld V2.1 (computer use) 79.6, slightly below two Claude models. Three usage gotchas: reasoning is always on (only low/high/max, can't turn off); needs a $1 top-up to unlock; every turn must resend full history.
One detail worth flagging: to go open source, the team actively removed cybersecurity data from mid-training and RL - an uncut checkpoint solves 90.5% on xdev-bench-hard pass@3, the released version only 81.0%. "Deliberately dialing down attack capability" will be a required lesson for every model going overseas.
First, the post-training arms race has taken over from pre-training - same parameters, better scores, proving there's still huge headroom in stacking post-training compute on a 3T base. Second, open-source price war has reached 1/5 of closed-source - performance hugging Opus 5.5 at 1/5 the price of GPT-6 Astra. Third, a cold splash: all scores are Moonshot's own, some on internal benchmarks; independent leaderboards haven't listed it. Wait a week for independent replication before migrating - the API is cheap, so the cost of trying is low.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| MiniMax M3.1 Flash open weights | 10-23 | MoE model (open-source drop) | At launch (10-19) only the API was open; this time the weights land on Hugging Face and ModelScope: 428B total / 23B active, 1M context, API $0.15/$0.60 per million. Vendor says 100T+ tokens processed in 72 hours; sparse attention + quantization-aware training + speculative decoding give ~3x the previous generation's inference speed and ~70% lower agent-task cost. Day-one ecosystem covers Huawei Ascend, T-Head, Kunlun, Moore Threads, NVIDIA, AMD - six chip vendors, with vLLM / SGLang / FlagOS support. The "day-one ecosystem package" is now standard for Chinese open models. |
| Seedance 2.5 4K (Volcano Engine) | 10-23 | Video generation (closed, enterprise trial) | AI video takes aim at the big screen: 10-bit color, 4K detail enhancement, better facial micro-expressions and reference consistency. The most practical bit is Draft mode - verify storyboards cheaply at 480P, iterate, then render the 4K master. Currently supplied to ~20 studios and directors worldwide, tied to specific projects, no public API. |
| RFD4-Proteina (Baker Lab x NVIDIA) | 10-23/24 | Generative biomolecular design (open source) | Nobel laureate David Baker's team with NVIDIA, on the BioNeMo framework: co-optimizing protein sequence, 3D structure, and interacting molecules in a unified "chemical-conformational latent space" (amino acids, sugars, DNA/RNA), with dual reasoning. Vendor-claimed 5-10x faster protein design, and first "de novo enzyme design" for active-ingredient synthesis. |
| Genie Studio 2.0 (Zhiyuan) | 10-23 | Embodied-AI platform | A one-stop embodied-AI platform: Flywheel model iteration + Agent orchestration + Monitor fleet ops, covering collect-train-test-infer-orchestrate-manage. Already in Changzhou CNC machine tools and Junpu BMS assembly lines, with vendor-claimed 99.9% precision-assembly success and 7x24 production. |
| Qiushi EIBench (CESI) | 10-23 | Embodied-AI benchmark system | A benchmark system plus five evaluation tools from the China Electronics Standardization Institute, covering data, algorithms, models, and robot systems - real-task performance, cross-scene adaptation, and safety. Fills a standardization gap in domestic embodied AI; read with Zhiyuan's platform: embodied AI is moving from "launch demos" to "verifiable delivery." |
| YOLO-Master (Tencent YouTu) | 10-24 | Real-time object detection (paper) | Accepted at CVPR 2026: the first deep integration of MoE into YOLO, with an ES-MoE module allocating compute by image complexity; YOLO-Master-N hits 42.4 mAP on MS COCO, balancing speed and accuracy. Open-source status TBD. |
Together, today's news is one line: Chinese open-source models are pressing on three fronts at once - performance in your face, price that crushes, compliance that leads. K3.1 Pro is 0.7 points from Opus 5.5 at 1/5 of Astra's price, with weights promised in a week. For model users the conclusion is direct: the same task's bill may be 20-50% of closed-source frontier, and migration cost has fallen low enough to re-run the math every quarter. The only thing left to wait for is independent replication - pretty vendor numbers only count once a third party has touched them.
(Data from official announcements and public reports, 2026-10-24; 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