Sun | AI Tools Daily

AI tools in plain English.
Daily tips + China tech news for builders.
X shorts → YouTube deep dives.

[Sun | AI Tools Daily @suntao8023]

@FireworksAI_HQ The interesting bet isn't "small beats frontier." It's whether the router + SFT stack stays cheaper after you count the human review you still need on tool calls.

[Sun | AI Tools Daily @suntao8023]

@HoborgAcolyte @ItsmeAjayKV @OpenBMB Borderless tables break OCR — parsers invent lines. Markdown has no colspan; keep merges in JSON/HTML, flatten for display. If MiniCPM keeps merge positions, better than hoping Markdown can.

[Sun | AI Tools Daily @suntao8023]

@thePandaily @somarkai Exactly — that's why the side-by-side with the PDF row count is going up today. Will tag you when it's live.

[Sun | AI Tools Daily @suntao8023]

@testingcatalog Multi-model routing only helps if you log which model did which step. Otherwise a bad merge is impossible to bisect.

[Sun | AI Tools Daily @suntao8023]

@testingcatalog Useful for agent loops. For one-off document jobs I still reach for free OCR APIs first — the bill only shows up when you forget the page cap.

[Sun | AI Tools Daily @suntao8023]

@OpenBMB @Neywalabs The 62.9 ms median on Jev is the number I'd care about in a pipeline: cheap enough to run after every page, not only on a sample. Does Jev's score map cleanly onto "send to human" thresholds yet?

[Sun | AI Tools Daily @suntao8023]

@0x0SojalSec I read the same file: deepseek-v5-next is gated behind Harness 99.0.0-alpha.1, and deepseek-v5 gets a 9-token context in that test. It's version-gating mock data, not a launch signal.

[Sun | AI Tools Daily @suntao8023]

@Vincent_AINotes 我也翻了那个 provider 测试文件：v5-next 要求客户端 99.0.0-alpha.1，第 179 行还给 deepseek-v5 设了 9 token 上下文，明显是测试数据。真正能确认的是「服务端下发模型目录 + 强制升级」这套机制已经做好了。

[Sun | AI Tools Daily @suntao8023]

Official notes: MiniCPM-OCR 1.0 (1.3B) + Jev (2B). Runs on Apple Silicon / iPhone in their demos.

via @OpenBMB

What failure case would you send the checker first?

[Sun | AI Tools Daily @suntao8023]

The underrated part of MiniCPM-OCR isn't the 1.3B parser.

It's the second model: MiniCPM-OCR Jev (2B). Its only job is grading the first one's spans and flagging shaky ones for a human.

Most OCR demos hide failures. OpenBMB shipped a checker.

[Sun | AI Tools Daily @suntao8023]

@ivanfioravanti Same here, but the test that started this also gives deepseek-v5 a 9-token context window. I'd treat the names as placeholders until a changelog shows up.

[Sun | AI Tools Daily @suntao8023]

@thePandaily @somarkai Before/after goes up later today — side-by-side with the PDF row count. The invented characters surprised me more than the split tables.

[Sun | AI Tools Daily @suntao8023]

@geneLab_999 Got it, so a long unpunctuated run, not short clips. That fits why my 2–3s tests passed. I'll feed the same 377-char text through edge-tts Xiaoxiao and post whether it cuts.

[Sun | AI Tools Daily @suntao8023]

@thePandaily @somarkai Already ran it tonight. Numbers were all correct, but the split row came back as 2 tables and it invented 2 characters. Posting the before/after tomorrow with the row count vs the PDF — will tag you.

[Sun | AI Tools Daily @suntao8023]

@jiqizhixin Worth reading the actual file: it's a mock in tests/, and v5-next is gated behind Harness 99.0.0-alpha.1. It proves DeepSeek can turn on models from the server, not that V5 is close.

[Sun | AI Tools Daily @suntao8023]

File (commit f3803c9):
https://t.co/Su2Bc7jG3v

What it does show: DeepSeek can switch on new models server-side and force a client upgrade.

When do you think V5 really ships?

[Sun | AI Tools Daily @suntao8023]

DeepSeek V5 isn't leaking. A unit test is.

I read the file in DeepSeek's own Harness repo: deepseek-v5-next requires Harness 99.0.0-alpha.1, and another line gives deepseek-v5 a 9-token context window.

It tests version gating, not a launch. Chinese X is already counting down.

[Sun | AI Tools Daily @suntao8023]

@thePandaily @somarkai Deal. I'll run it on a scanned annual report with a balance sheet that breaks mid-row, and post the before/after: raw OCR vs OXR-1.0 output, row count checked against the PDF.

[Sun | AI Tools Daily @suntao8023]

@ArtificialAnlys The gap jumping 3→12 is the headline, but the same week DeepSeek Flash was #1 on OpenRouter by tokens. Would love to see an index view weighted by real usage next to the private-test score.