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.

The recruiting company BOSS Zhipin quietly shipped its lab's biggest model yet: Nanbeige4.3-33B. Weights hit Hugging Face on Oct 24, officially announced the next day, under Apache 2.0 - free for commercial use - with the Base version released too, so others can keep training on it.
Positioning in plain terms: it's not a chat model, it's a coding-agent model built for "AI doing work" - running terminal commands, editing codebases, office automation, that kind of long-chain task. The lab had only shipped 3B models before; this 33B is its first "big" one.
Three lost items: SWE-Bench Pro 60.3 vs 61.7 (slight loss to Qwen3.8-27B, the hard real-repo bug-fixing benchmark), GDPval 84.8 vs 86.4, GPQA 89.5 vs Gemma4-31B's 91.3. Two bigger caveats: first, all scores are Nanbeige's own - independent leaderboards haven't scored it; second, outside reviewers found some competitor scores in its comparison table don't match official model cards (suspected harness differences or even row/column swaps). Until third parties replicate, treat the comparison table as "directional." Also, no inference vendor hosts it yet - self-host or wait for our plaza to add it.
First, even "big-company side projects" are competing in open source - a recruiting company ships a 33B coding model, showing the bar has dropped to where a business's AI team can mass-produce a flagship-class small model. Second, "teacher first, student later" is the new open-source play - skip the trillion-parameter hype, train a mid-size to be competitive, then distill to a 3B edge model. For enterprises wanting private deployment, this route beats chasing the biggest cup.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| Carbon-X (HF Bio x ZGC x Federico II) | 10-25 | DNA language model (open source) | The first 14.8B open model putting "DNA sequence" and "English questions" in one model (Qwen3-14B base, Apache 2.0): predicts whether a mutation is pathogenic, marks regulatory regions, 66.7% average across 5 DNA benchmarks (self-reported), just 0.8 below GPT-6.1 Sol; runs on one H200, with vendor-estimated inference cost ~1/1800 of frontier models. Can also serve as a "subagent" lifting Kimi K3's DNA index from 49.4% to 60.5%. Note: input cap ~50K base pairs, can't take a whole genome, no wet-lab or clinical validation - research aid only. |
| TokenGate (Stanford x UC Berkeley) | 10-25 | Long-text acceleration (open source) | Adds a trainable "skimming switch" to existing models: score input tokens, let only the critical ones into full compute. Report: 1.4-11.3x faster first-token latency, 1.5-3.4x lower peak memory with no accuracy loss; a module trained at 16K still speeds up 13x at 128K. Apache 2.0, code and weights open. The long-context cost/latency problem is being nibbled by "don't change the model, add a component." |
| Microsoft Agensh officially open-sourced | 10-24 | Multi-agent framework (MIT) | The paper circulated earlier (flagged as "old-paper spread" in our 10-25 issue); the code is now open. Counter-intuitive design: no central scheduler - 1,024 agents claim tasks, cross-verify, and self-merge in a shared repo. On the hardest coding tasks, GPT-5.6-sol's pass rate rose from 19.31% (single agent) to 28.78% (128 agents); 55.06% on pandoc with 1,024 agents. Note: agent count isn't cost-effectiveness - token and coordination costs are unaudited. |
| MCP reference servers v2.1.0 | 10-25 | Agent tool standard (security fixes) | Fixes 8 security issues across the filesystem, Everything, and Fetch reference servers: directory-access bounds, web-origin validation, network-request checks. Teams handing local directories or network access to tools should upgrade soon - tool-layer bugs are an agent's "back door." |
| pplx-decider-v1.2 + flash decision models (Perplexity) | 10-24/25 | Decision models (open source) | After OpenAI opened its Decisions API and Ant shipped LingDT, Perplexity also open-sourced its lightweight decision model. Decision/routing "small-fast-agile" tasks are becoming another crowded open-source niche after coding (no detailed scores; capability boundary awaits third-party verification). |
No mega-flagship this weekend; the theme is "specialization + going downmarket." A 33B coding teacher, a 14B DNA specialist, a long-text accelerator - none chase "the strongest," all chase "good enough, cheap enough, single-card runnable." The open-source battleground has moved from trillion-parameter launches to 24GB GPUs. The practical takeaway for developers: if your task is coding agents or vertical Q&A, check sub-33B specialized open models first - likely good enough, at a fraction of the frontier bill.
(Data from official announcements and public reports, 2026-10-26; 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