Open Models

MiniMax M3.1 Flash, Zhipu's GLM-5.3-Flash and DeepSeek's V4.1-Flash-1017 at off-peak rates all list input at 15 cents a million tokens. Cached input, most of an agent's bill, runs from 0.3 cents at DeepSeek off-peak to 3 cents at Zhipu. Third-party hosts already serve DeepSeek's and Zhipu's models for less; MiniMax has not released M3.1 Flash's weights, so only MiniMax can sell it.

On 3 October 2026 the three AI-made films in iQiyi's 《灵魂摆渡》 浮生梦 series crossed 10 million yuan in cumulative revenue-share box office, about 1.49 million dollars at the onshore rate of 6.7050, the last print before the National Day closure. The figures are iQiyi and 长信传媒's own, carried by leiphone on 3 October. 分账票房 is not ticket sales and the English phrase will mislead: a licensed web film earns a per-view cut set by the platform's tier schedule, and the announced figure is the producers' share after the platform takes the rest, so the 10 million is money that reached 爱奇艺 and 长信传媒 rather than a gross to be divided afterwards. The sequencing carries the claim. The first two films went up exclusively on iQiyi on 22 August and recovered the production cost of both inside their first week; first-month cumulative revenue share passed 8 million yuan, about 1.19 million dollars; the third launched on 1 October and took the series past ten. The second number in the same announcement is the one nobody quotes: across the three films the production accumulated nearly 500 character assets, nearly 800 scene assets and nearly 200 core prop assets, all of which iQiyi describes as reusable in later work, built with what it calls AI 技术 plus 真人表演迁移, live-action performance transfer, over a legacy television property from the early 2010s with an audience that already remembers specific faces. The regulatory half ran the same week and took no English notice at all. 《后西游记》, billed by ithome as China's first AIGC long-form drama, began its 战天宫篇 arc on 4 October on Hunan Satellite TV and Mango TV; season one runs 30 episodes at 40 minutes each, and the first arc, 花果山篇, aired five episodes from 31 August in Hunan's primetime slot and placed first in real-time ratings among provincial satellite channels in that hour. ithome describes it as the first series to use 边审边播, review while broadcasting, after the 广电21条, the arrangement where a show is produced, cleared and broadcast concurrently rather than in sequence. Set that against the model layer: Artificial Analysis's AA-Video-T2V v2.0 board, scored on blind votes from a recruited human panel, reads Wan 3.0 first at 1156 Elo, Utopai X at 1149, Dreamina Seedance 2.5 at 1142 and MiniMax H3 at 1138, and Utopai Studios is a California company whose model is a post-train of MiniMax H3. MiniMax H3 also leads the open-weights field on that board, 192 Elo points clear of LTX-2.5 Fast at 946.

DeepSeek's newest open-weight release is smaller on disk than its most famous one, and that decides a procurement question in China. Read live off the Hugging Face API, DeepSeek-V4.1-Flash ships as 48 safetensors shards totalling 510,296,708,312 bytes, or 475.3 GiB, against DeepSeek-R1's 163 shards and 688,586,727,753 bytes, or 641.3 GiB: a difference of 166.0 GiB, 25.9 percent of the older file. The reason sits in config.json, where the experts are declared fp4 and the remaining weights fp8, in 32-by-32 weight blocks with ue8m0 scales, so DeepSeek shipped the quantisation itself instead of leaving a buyer to quantise a BF16 release and absorb what that costs in quality. The model card puts the global KV cache at 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash, which binds harder than the weights do once several people use the machine at once. 钛媒体 ran a piece by 半导体产业纵横 on 3 October 2026 that reaches the same conclusion from the buyer's side: more than sixty vendors brought a 大模型一体机, the all-in-one LLM appliance, to market after R1, a full-specification unit ran 1.2 to 1.5 million yuan, about 179,000 to 224,000 dollars at the onshore rate of 6.7050, and builds on domestic accelerators were often quoted two chassis and a budget near 4 million yuan, around 597,000 dollars. The market cooled within three months on a consensus the piece quotes flatly, that DeepSeek is useful and DeepSeek appliances are not: eight cards in one box get no cluster economics, and by the second half of 2025 the full-specification tier had stopped selling while the volume at smaller vendors moved to 32B and 70B models. The piece adds that 2026 price rises in memory, SSDs and GPUs have gone far enough that a unit bought in early 2025 can recover its entire purchase cost on the second-hand market, and carries an anonymous account of five Ascend-card servers sold to a state enterprise R&D department for just over 10 million yuan, about 1.49 million dollars, then resold at a profit to a neighbouring team after the project failed and the team was dissolved. That account names nobody and cannot be verified, so read it as the story the industry is telling itself. One caution on the parameter arithmetic, which the issue makes itself: the card calls V4.1-Flash a multimodal mixture-of-experts model with a 552B backbone against R1's 671B total, but it also lists Engram conditional memory at 196B parameters sparsely accessed by token lookup, and 552 plus 196 is more than 671, while Hugging Face's own tensor census for the repository comes to 763 billion elements. The parameter comparison and the file comparison are two different arguments that happen to agree, and only the file one answers whether the weights fit in a box somebody already owns.

Meituan launched LongCat-2.5-Preview on 25 September on its LongCat API open platform, and the release notes list three things it does: it understands images, it is good at code, and it is deeply adapted to Claude Code and other mainstream development environments. That third bullet has been on every LongCat release since 20 April 2026. The setup guide repoints ANTHROPIC_BASE_URL at api.longcat.chat/anthropic and maps four Claude Code model variables, including the ones named for the Opus and Sonnet tiers, to LongCat-2.5-Preview; the platform documents itself as fully compatible with the Anthropic Claude API format and ships thirteen such client configurations. The pricing page carries two yuan per million input tokens on a cache miss, four fen on a cache hit and eight yuan per million output, with a dollar panel reading thirty cents, six tenths of a cent and $1.20, and those two panels reconcile at Friday's onshore rate of 6.7125. The predecessor LongCat-2.0's pricing page carries the identical three numbers, so nothing is charged for the size, the image understanding or the coding work. The 1.6-trillion-total, roughly 48-billion-activated and million-token-context figures are not on any Meituan primary and are attributed to IT之家 citing the platform's release notes. The open-weights record is the specific part: Meituan has published twenty-nine models on Hugging Face including LongCat-2.0's weights on 5 July, has published nothing since 31 July, and LongCat-2.5-Preview has no weights at all. The download counts complicate the obvious reading, with the trillion-parameter LongCat-2.0 taken 1,773 times in thirty days against 23,795 for the much smaller LongCat-Flash-Chat.

Harvey, the OpenAI-backed legal AI company, published its first post-trained model on 20 August. Harvey Tenet is a Kimi K3 base, post-trained with a rank-64 LoRA over the full network on a dataset of about 1,750 tasks. It completes almost twice as many held-out tasks on the Legal Agent Benchmark as base Kimi K3, and places second on that benchmark overall. Harvey ran ablations on candidate judge models against heavier frontier models and settled on Kimi 2.6 as the grader. Its M&A diligence work uses a GLM-5.2 orchestrator, taking rubric criteria pass rate from a 43.8 percent best baseline to 60.1 percent after self-distillation post-training.

Alibaba’s Qwen now accounts for more than half of every open-source AI model download on earth and has passed a billion downloads on Hugging Face, displacing Meta’s Llama, per SCMP and Pandaily, though those share figures trace to a single report. The lead is firm enough that Alibaba plans to start charging heavy users of its next open model, and in the same week Meta and Nvidia both re-entered the open-weight market they had backed away from, Meta with Muse Glimmer and Nvidia with Nemotron 3.5 Lightning, explicitly to catch the Chinese labs now setting the open-source default.

DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.

An IDC survey found 47 percent of large US firms now run a Chinese model in at least one use case, and at one point mid-July all five top models on OpenRouter were Chinese. Coinbase, DoorDash, Airbnb and Cursor are named adopters, switching to cut the bill as UBS puts Chinese API prices at 10 to 20 percent of the US alternative.

DeepSeek open-sourced an 85% inference speedup. OpenAI found the same trick the same fortnight and kept it secret. One bets on standards, the other on margin.

DeepSeek open-sourced V4 under MIT with Huawei Ascend as a primary deployment path: 1.6 trillion parameters, million-token context, coding a hair behind the best Western models. The first sovereign stack.

A team completed what it says is the first third-party full-parameter post-training of a 1.6-trillion-parameter model on about 1,000 Huawei Ascend cards, with no instability. The penalty is efficiency, and it is shrinking.