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.

iFlytek unveiled Spark X2.6 at its 1024 Developer Festival in Beijing. The hook isn't the biggest parameter count - it's one sentence: this is the first flagship-class model claimed to be trained entirely on domestic chips.
In plain terms: Chinese models have mostly trained on NVIDIA cards, so a supply cutoff would stall the whole pipeline. X2.6 aims to prove you can train a 293B flagship to competitive levels on domestic chips. The significance isn't one model - it's that "this road is viable."
Domestic chips have long been memory- and bandwidth-limited - fine for short chats, but they crawl on long context and long-horizon tasks. Two numbers worth remembering:
First, all scores are self-reported - no independent lab has replicated them, and "SWE first" only holds for some sub-items (Verified actually loses by 0.2). Second, cloud weights aren't open - no private deployment, API or their bundled plan only. Third, "the only one trained on all-domestic compute" is a vendor claim about flagship scale; smaller models have run on domestic cards for a while. Before migrating, run your own real workload, don't just read leaderboards.
A financial footnote: iFlytek's first three quarters revenue was ¥18.187B (+7.05%), with net profit turning positive, but non-GAAP net loss was still ¥763M; Q3 R&D alone exceeded ¥1.7B, over 25% of revenue. The domestic-compute road is being paved with real money.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| Recursive Flint (Flint-40B-A2B) | 10-21 | Open weights (Apache 2.0) | The first model from Recursive (co-founded by Yuandong Tian and Richard Socher), built for local coding agents. 39.8B total / ~2.5B active per token, 6 of 256 routed experts + 1 shared, 262K context. Vendor-measured: 131 token/s on M4 Pro (48GB), 155 token/s on RTX Spark, ~5x same-class dense models. Six coding benchmarks average 80.6 (Qwen3.8 is 84.7). Trained from tokenizer to RL in under 4 months; its in-house auto-research system Eureka Machine handled architecture search, quantization (71% smaller), and GPU kernel optimization. INT4 ~23GB, runs in 48GB. |
| Claude Duet | 10-21 | Proactive agent app (research preview) | Based on Claude Opus 5.5, on iOS / Android / web, waitlisted for US Pro/Max users. Unlike a chatbot it carries its own VM and browser, keeps working when your device is off: pushes a morning brief, watches connected apps during the day, messages you only when a decision is needed. 6000+ connectors (Gmail, Slack, Shopify...); sending, paying, signing, and deleting all require your confirmation; never touches passwords or login tokens. Anthropic admits the preview can misjudge urgency. |
| MWA-Phoenix (open source) | 10-22 | Embodied world model (7B) | An on-device 7B latent-space world model with full pre-training, topping LIBERO-Plus, LIBERO-Pro, RoboCasa GR1, Bench2Dex, RoboTwin 2.0, and RoboDojo all at once (sim, real robot, real scene). No depth sensors, no agent dependency, no test-data replication; Phoenix-ILA folds motion and contact into action learning, and Phoenix-PhysCanvas lets it self-evolve in the cloud before deployment. Code and weights open. |
| Tev-27B / Tev-320B (UCLA x Adobe) | 10-22 | Long-horizon agent training set + open models | Targets the "open models can't do long work" gap with the first long-horizon agent training set carrying verifiable dense rewards - Tev-Gym (1,170 environments, 7 domains), plus PPRS reward shaping. Tev-27B hits Pass@5 0.28 on ALE-CM, on par with GPT-6 Astra and Claude Opus 5; Tev-320B is the strongest open model on that Terminal-Bench 4.0 subset. Paper public (arXiv 2610.18868). |
| GameScaling (miHoYo + Tsinghua/Peking U. et al., open source) | 10-22 | Game-operation model (0.8B-27B) | From the Honkai AI team with Tsinghua, PKU, UCAS, HKU, NUS; continued pre-training on Qwen3.5 VLM over keyboard-mouse recordings. One call outputs ~200ms of actions (6 steps x 33ms + mouse move and scroll). Covers Minecraft mining/crafting/combat, Doom combat and pathfinding, escape-room puzzles. 8 checkpoints: 0.8B/2B/9B/27B in two builds each (S1 Minecraft-only, S3 multi-game 27%), 1,600 hours each. Apache 2.0, vLLM; the 27B needs 2-card tensor parallel. |
| Mirendil-Base | 10-21 | Code base (100B-500B, not open) | One researcher + a team of agents, 3.5 months, under $10M total, pre-training a 100B-to-500B code base from scratch on public data alone. The more striking part: 99% of the pre-training code was written by AI agents. Next up: Base-Large (500B-2T) and Base-XL (2T-10T). Independent evals, post-training scores, and API in weeks - no comparable benchmark yet; treat as a direction signal. |
| SparkAudio 1.0 (iFlytek) | 10-22 | Voice foundation model | iFlytek's first voice base, folding translation, recognition, semantic understanding, and emotion recognition into one model, with arbitrary Chinese-English switching (not a serial recognize-then-translate flow). Ships alongside Spark X2.6 - the "one model eats the whole voice chain" route. |
| Kling 4.0 API | 10-21 | Video generation API (commercial) | The model itself went GA on Oct 15; the API premiered the evening of Oct 21, rolling out to enterprise customers and developers - the real "usable day" for anyone wiring video generation into a workflow. 3-30s per shot, up to 10 keyframes, 15 multimodal references (10 images + 5 videos), dual-channel stereo, 9-language lip-sync, 8000-token prompt cap, 21:9 ultrawide. Note: 1080p/4K 10-bit HDR is still "coming soon" - don't count it as shipped. |
Every story today is really one story: the cost structure of building models is being rewritten. X2.6 proves domestic chips can train a flagship at ¥0.9/M, Flint ships a 40B model that runs 130 tokens/s on a laptop in 4 months, and DeepSeek raises 100B yuan to push past 2T parameters - different directions, same conclusion: "how many cards you have" is no longer the moat, "how fast you turn ideas into a usable model" is.
(Data from official announcements and public reports, 2026-10-22; 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