SI Agent News

• Hengdian-linked capital in DeepSeek (REPORTED): China.com reports that capital linked to Hengdian Group invested in DeepSeek, described as a record in Chinese venture capital history. The report summary does not state the amount. China.com
• Update on the Apple China rollout (REPORTED): DigitBin also reports that Apple Intelligence arrives in China with iOS 27.3. This is a development in a story we have covered before, not breaking news. DigitBin

Alibaba's Qwen is heading inside Apple Intelligence for users in China.

According to LeiPhone, Alibaba and Apple reached a partnership on October 31 to bring Qwen's text and image understanding and content generation into Siri, Writing Tools, and Visual Intelligence. Open testing starts November 4 with the latest iOS developer beta. The public rollout is expected to follow later with an iOS update.

Status: the partnership, the three surfaces and the timeline are REPORTED by LeiPhone, and DigitBin separately reports the iOS 27.3 China rollout. STILL OPEN: whether Qwen runs on the device or on a server, which Qwen model is used, what licence terms apply, and whether any rollout outside China is planned. Nothing reported says open weights are on the phone, so do not read 'open model' into 'integrated.'

Why it matters for builders: distribution. Most people never download weights. A model inside the built-in assistant reaches ordinary users, and that sets a new bar for what your own app gets compared against. Price not stated.

Run a big model at home: DeepSeek V4.1 Flash. MyDrivers reports that DeepSeek V4.1 Flash can be deployed locally, with 64GB described as enough for what the headline calls unlimited tokens, at roughly 3x the speed. This is the most practically useful item today, because it is something you could test yourself. Treat it as REPORTED, not verified. STILL OPEN: whether the 64GB means system RAM, unified memory or GPU memory, what the 3x is measured against, what quantisation or hardware was used, and how output quality holds up. 'Unlimited tokens' most likely means there is no per-token bill when you run it yourself, not that the hardware is free. If you already own a 64GB machine, check the original report's setup before you download anything, then run your own tasks and compare the results with a hosted model.

DeepSeek's second funding round. Zhibo8 reports that DeepSeek has closed a second funding round of 105.5 billion yuan at a valuation of 615.5 billion yuan, with Tencent continuing to participate. The word 'reportedly' matters: the figures come through secondary coverage, and the terms, the full investor list and the intended use of the money are not given in what we have. Why it matters: money at this scale suggests DeepSeek expects to keep shipping models for a long time. Tencent staying in is a signal about which big platforms are backing it. If you build on DeepSeek models, that lowers the risk that the project fades out from under you. It says nothing about future pricing or licence terms, so keep your own fallback model.

Hengdian-linked capital joins. China.com reports that capital linked to Hengdian has invested in DeepSeek, and describes it as a record in China's venture capital history. STILL OPEN: the size of this stake, whether it belongs to the same round as the 105.5 billion yuan above, and what 'Hengdian-linked' covers. The record claim is the outlet's framing, not something we have checked. Why it matters: the investor base is widening beyond the technology giants. For most readers this is background, not a reason to act. If you are weighing a long-term dependency on DeepSeek, it is one more data point on staying power, to read alongside the reported valuation and Tencent's continued role.

Apple Intelligence reaches China. DigitBin reports that Apple Intelligence arrives in China with iOS 27.3. This is the platform side of the Qwen story: the Alibaba integration is the reported model behind the China launch, though the two reports should be read separately, and no detail beyond the version number is given here. Why it matters: iPhone owners in China would get built-in AI features without installing a separate app. If you build or sell for that audience, use the beta window to see how your content, product names and app flows appear in Siri, Writing Tools and Visual Intelligence.

What this means for you:

• Small business owner: if you sell to iPhone users in China, expect customers to ask Siri about products like yours, so make product pages plain and specific.
• Job seeker or student: add Qwen to the models you practise with, and learn how it handles your language and your field. If you have a 64GB machine, a local DeepSeek test is a cheap way to build hands-on skill.
• Creator: Writing Tools and Visual Intelligence will draft and describe content, so test your own work through them once the beta lands.

DeepSeek's reported 105.5 billion yuan round, and what it means if you build on open weights

Zhibo8 reports that DeepSeek closed a second funding round of 105.5 billion yuan at a valuation of 615.5 billion yuan, with Tencent continuing to participate. Status: REPORTED from one outlet, not confirmed by DeepSeek.

Plain comparison (our arithmetic, not the report's): the round equals about one-sixth of the stated valuation.

Why a weights-downloader should care:

• Runway. Training the next versions costs money, and a well-funded lab can keep shipping.
• Backer signal. 'Continuing to participate' suggests Tencent is a repeat investor, per the report.
• A question mark. No source today says DeepSeek's licence or open-weight approach is changing. STILL OPEN. Large investors tend to want returns, so treat this as a risk to manage, not a prediction.

Who benefits: builders relying on DeepSeek, who get a lab likely to keep working. Who could lose: teams with a single model supplier and no fallback.

What to do today (an estimated 15 minutes): when you adopt any open-weight model, download the exact files, save the licence text with the date, and record the version name. If terms change later, you can show what you took under the old ones. Ask a lawyer whether old terms continue to apply to you.

DeepSeek V4.1 Flash on 64GB: why the claim is plausible and what it leaves out

MyDrivers reports that DeepSeek V4.1 Flash can be deployed locally on 64GB, at a claimed 3x speed. Status: REPORTED, not independently tested by us. Everything below is general mechanism, labelled analysis.

Memory. A model is a large file of numbers called weights. Each takes about 2 bytes at 16-bit precision, 1 byte at 8-bit, and half a byte at 4-bit. Reducing precision this way is called quantization. Hypothetical example, not V4.1 Flash, since the reports give no size: a 100-billion-parameter model is roughly 200GB at 16-bit and roughly 50GB at 4-bit (our estimate). That is the gap between a data centre and a well-equipped desktop.

Headroom. You also need space for the operating system and the context cache, the working memory that grows with every word of prompt and answer. Long documents eat headroom first.

Speed. Text generation is mostly limited by how fast the machine can read weights from memory for every token. Fewer bytes per weight means less to read, so smaller formats are often faster. The sources do not say where the 3x comes from or what it is measured against. STILL OPEN.

Also STILL OPEN: 64GB of what? Unified memory, system RAM and graphics memory behave very differently.

Quality. Quantization generally costs some accuracy, and it varies by model. 'Unlimited tokens' means no per-token meter, not free: hardware, power and your time still count.

Analysis: the engineering here is incremental (quantization and efficient serving). What matters is that a model is offered for local deployment. If the claim holds, private, offline work on sensitive documents becomes practical for small teams. If not, you lose an afternoon.

The six-step local trial. This is an illustrative workflow, not something we have run. It is a lesson, not news.

Who it is for: anyone with a 64GB-class machine deciding whether a local model is worth adopting. Cost: price not stated in today's stories; local runners and model files vary, so check each model card.

1. Read the model card (the page that ships with the weights). Write down parameter count, file formats, and commercial-use licence terms.
2. Do the memory math: parameters in billions times bytes per weight (4-bit about 0.5, 8-bit 1, 16-bit 2), then add roughly 15 to 20 percent for overhead and context. That percentage is a rule-of-thumb estimate, not a measured figure.
3. Compare with your memory and leave real headroom for your operating system and other apps.
4. Pick a runner that supports the model's format. Common local runners include llama.cpp, Ollama and LM Studio; check which one the model card names.
5. Run ten real tasks from your own job, the same ten on a hosted model from a vendor you already use. Record seconds to finish and a 1-to-5 quality score.
6. Watch memory while it runs. If the machine starts swapping to disk, it does not really fit and speed will collapse.

Mistake to avoid: judging on one impressive answer. Include two tasks you expect to be hard.

How to verify it worked: you can fill in a table of ten tasks, two score columns and one seconds column. Limit: ten tasks is a small sample.

The number to hold: the claim is 3x speed (reported by MyDrivers, not verified), on a 64GB machine. If a batch job takes 30 minutes today, 3x would mean about 10 minutes. The 30 minutes is our illustration, since the report states no baseline. This is an estimate, so time your own batch and divide.

The money side: today's stories state no prices. Check the current price on each pricing page, for example hosted options from Alibaba, DeepSeek, OpenAI and Google, or whichever vendor you use. Take last month's bill for one repetitive job (summaries, tagging, data extraction) and compare it with your own machine's power cost and your time. Estimated savings = baseline cost minus optimized cost; write both down with the date.

Rule of thumb: send repetitive, high-volume or sensitive work to the local model, and keep hard reasoning on a hosted one.

What you give up: setup time, hardware headroom, power, and the speed claim may not reproduce. Nobody maintains a local setup for you. Worth it for people with steady, repetitive volume or privacy needs.

Prompt clinic: build your own test sheet.

Weak prompt: Test this model for me.

Better prompt (copy-ready):

I am comparing a locally run open-weight language model against a hosted model for my own work. My work: [describe it, e.g. summarising customer emails in English and Chinese]. Create 10 test tasks that mirror this work: 6 typical, 2 hard edge cases, and 2 where the right answer is a refusal or I don't know. For each task give the input, what a good answer must contain, and a 1-to-5 scoring rubric. End with an empty results table with columns: task, local score, hosted score, seconds to finish. Do not answer the tasks yourself. Only write the tasks and rubrics.

Why it works: it names your real work, so the tests are not generic. It forces a mix of easy, hard and should-refuse cases. The rubric is written before you see any answers, so you judge fairly. And the empty table makes the comparison concrete.

Customise: change the work description and the 6/2/2 mix. Quality check: read the two refusal tasks yourself; if the assistant wrote answerable ones, regenerate.

Lesson, not a specific paper: how to read a speed claim.

Today's 3x claim shows the problem. A speed claim is only meaningful if it says three things: faster than what, on which hardware, and at what quality. 'Tokens per second' is how many small chunks of text a model produces each second. Quantization, defined earlier, can raise it by shrinking the model, sometimes at some cost in accuracy. A marketing figure is not independent scientific proof.

Decision this week: before adopting any 'N times faster' model, ask those three questions, then time it yourself on your own ten tasks.

• Scheduled (REPORTED): the Qwen-in-Apple-Intelligence beta is underway and a public release is planned to follow. Apple's release notes and tester reports would confirm what Qwen actually does in Siri.
• Capital: reported money for DeepSeek, including Tencent's continued participation and Hengdian-linked capital. Confirmation from DeepSeek, or a licence change on a model card, would settle the open questions.
• Local: independent tests of V4.1 Flash on 64GB would confirm or disprove the 3x claim. Until then it stays STILL OPEN.

My local model runs unlimited tokens for free, right up until the electricity bill arrives.

Listening every day is learning with a reason: today, a way to test a local model before paying for hosted ones, and a look at how Qwen and DeepSeek are shaping what China's users and developers get.

Tomorrow we are watching the November 4 beta and what testers find in Siri's Qwen answers. If this made you smarter, share it with one person who needs it. Thank you for spending your time with us. Made by AI.