Junyang Lin(@JustinLin610)

one thought about possible misunderstanding: here we are discussing more about strategic use of mode...

Junyang Lin (@JustinLin610) — Thu, 16 Jul 2026 03:47:58 GMT

one thought about possible misunderstanding: here we are discussing more about strategic use of models instead of technical methods. we did things like imagining moe really playing like experts or setting up a task router to save costs but found them hard to work well. here what I am seeing is more about a rough standard of choosing different models. q and a and very simple agentic tasks like problems solved by using tools once or twice or fixing a small bugs, compared with building a large code projects or doing ml experiments and evals, etc.

💬4🔄1❤️14👀3264📊6

⚡ Powered by xgo.ing

nice work and great to see new arch working! btw the small model is not opensourced?

Junyang Lin (@JustinLin610) — Thu, 16 Jul 2026 03:42:03 GMT

nice work and great to see new arch working! 

btw the small model is not opensourced?

Thinking Machines@thinkymachines

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

thinkingmachines.ai/news/introduci…

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

🔗 View Quoted Tweet

💬4🔄0❤️99👀21581📊13

⚡ Powered by xgo.ing

i recently discussed with some friends about the strike of model effects and costs. while frontier m...

Junyang Lin (@JustinLin610) — Tue, 14 Jul 2026 11:21:57 GMT

i recently discussed with some friends about the strike of model effects and costs. while frontier models are smashing a lot of things it is always expensive to use them. me personally always think about using as many tokens as possible with my subscriptions and coding plans, and i think if i turn the usages to tokens that might be a much bigger amount of money (i remember a media previously said if u use codex crazily u can use over 8000 dollars a month with 200 dollar subscription. not sure about the number). 

how about enterprises that really use tokens and count expense on the unit of token? some interviewees told me that they tried their design their system by tasks, like easier or to say more formatted tasks go to seed, qwen, dpsk, harder tasks like agentic tasks go to fable or sol. glm 5.2 is nice for coding but still after claude and gpt, recent grok is a bit similar. 

this just reminds me that if we design the system well on the user side, we can have more productivity based on sufficient usage of ai today without too much worry about expense. maybe there is still much room for this.

💬23🔄6❤️110👀22563📊40

⚡ Powered by xgo.ing

anything interesting recently in the opensource community?

Junyang Lin (@JustinLin610) — Mon, 13 Jul 2026 04:04:59 GMT

anything interesting recently in the opensource community?

💬18🔄0❤️95👀24339📊29

⚡ Powered by xgo.ing

cool

Junyang Lin (@JustinLin610) — Fri, 03 Jul 2026 10:49:58 GMT

cool

Mira Murati@miramurati

Bridgewater used their unique financial knowledge and partnered with us on @tinkerapi to fine-tune a model that helps their analysts focus on what's important. Experts improving AI that empowers experts.
thinkingmachines.ai/news/learning-…

🔗 View Quoted Tweet

💬1🔄1❤️37👀20590📊5

⚡ Powered by xgo.ing

jb in exchange of pg? what?

Junyang Lin (@JustinLin610) — Fri, 03 Jul 2026 05:47:17 GMT

jb in exchange of pg? what?

💬6🔄0❤️29👀19034📊10

⚡ Powered by xgo.ing

wen people say sd they mean seeddance instead of stable diffusion. f hard for me to react on this

Junyang Lin (@JustinLin610) — Sun, 28 Jun 2026 09:35:20 GMT

wen people say sd they mean seeddance instead of stable diffusion. f hard for me to react on this

💬11🔄1❤️76👀19817📊20

⚡ Powered by xgo.ing

i still remember the discussion of the case of us visa application and i thought that must be missio...

Junyang Lin (@JustinLin610) — Fri, 26 Jun 2026 15:21:40 GMT

i still remember the discussion of the case of us visa application and i thought that must be mission impossible for ai and just said u guys must be crazy... now it seems that i am the dumb hahah!

anyway, time for frontier models to fight again! bon courage!

XLANG NLP Lab@XLangNLP

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved?

Not even close.

🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks.

What's new:
🎯 108 real-world workflows, each ~1.6 hours ⏱️ for a skilled human
⚙️ ~318 tool calls/task vs. ~30 in OSWorld 1.0
🌍 Grounded in authentic artifacts & stateful user profiles
⚡ Captures real phenomena: dynamic environments, streaming interaction, cross-source reasoning, implicit-state inference & more

📊 Best results: Claude Opus 4.8 reaches the highest accuracy at 20.6%, while GPT-5.5 is far more token-efficient but plateaus near 13%. No one is close to solving real computer use.

🏠 Homeposworld-v2.xlang.ai0pFwC
📄 Pgithub.com/xlang-ai/OSWor…TOiR1t
💻github.com/xlang-ai/OSWor…F8CoqgN
🤗 Dhuggingface.co/datasets/xlang…HHHygFk7

🧵 [1/8]

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬3🔄8❤️159👀37908📊26

⚡ Powered by xgo.ing

y codex so dumb these several days

Junyang Lin (@JustinLin610) — Sat, 20 Jun 2026 15:22:20 GMT

y codex so dumb these several days

💬23🔄2❤️102👀91825📊36

⚡ Powered by xgo.ing

exactement

Junyang Lin (@JustinLin610) — Wed, 10 Jun 2026 08:21:05 GMT

exactement

Jeremy Howard@jeremyphoward

@karpathy This is not a day for celebrating, Andrej.

It's a very dark and very sad day, and the damage may be impossible to undo.

🔗 View Quoted Tweet

💬7🔄1❤️93👀25353📊14

⚡ Powered by xgo.ing

ai making ai better recursively is an impressive idea but sometimes depressing as human researchers ...

Junyang Lin (@JustinLin610) — Sun, 07 Jun 2026 07:10:19 GMT

ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.

💬21🔄13❤️222👀44740📊57

⚡ Powered by xgo.ing

what do u think of gemma 4 12b?

Junyang Lin (@JustinLin610) — Thu, 04 Jun 2026 17:04:36 GMT

what do u think of gemma 4 12b?

💬32🔄2❤️100👀36696📊44

⚡ Powered by xgo.ing

why preview

Junyang Lin (@JustinLin610) — Fri, 24 Apr 2026 07:24:53 GMT

why preview

DeepSeek@deepseek_ai

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.

Try it now chat.deepseek.com1Dl via Expert Mode / Instant Mode. API is updated & available today!

📄 Tech Repohuggingface.co/deepseek-ai/De…kYtp
🤗 Open Weighuggingface.co/collections/de…i7SDM

1/n

🔗 View Quoted Tweet

💬31🔄4❤️361👀95836📊59

⚡ Powered by xgo.ing

i do like this passage, and here are some thoughts: 1. critical thinking is essential in the era of...

Junyang Lin (@JustinLin610) — Mon, 13 Apr 2026 17:54:58 GMT

i do like this passage, and here are some thoughts:

1. critical thinking is essential in the era of agents. i still remember that many years ago when i studied the lesson of critical thinking, i learned that keeping debating with yourself by listing out reasons can really deepen your thinking. today, critical thinking becomes humans debating with agents, so that they can think more deeply together and analyze problems in a more comprehensive way.

2. designing a healthy and well-structured organization and system is essential for creation and building. with systematic support and efficient tooling, humans can work exponentially more effectively together with agents. that gives people more time to take care of their physical and mental health, while also exploring new opportunities.

3. new era often favors newbies, because they have less past experience and therefore less fear of current difficulties. what oldbies should really think about is which parts of their experience are actually worth leveraging. from my perspective, we should think more carefully about which experiences are truly aligned with first principles.

but anyway, ai first is super, super exciting!

Peter Pang@intuitiveml

x.com/i/article/2043…

🔗 View Quoted Tweet

💬11🔄25❤️273👀70248📊75

⚡ Powered by xgo.ing

we need agent evals that are really consistent with real world usages. otherwise people are optimizi...

Junyang Lin (@JustinLin610) — Sun, 12 Apr 2026 00:36:31 GMT

we need agent evals that are really consistent with real world usages. otherwise people are optimizing foundation models for the wrong direction. the problem of targeting is even bigger than benchmaxxing.

💬22🔄20❤️254👀33204📊63

⚡ Powered by xgo.ing

happy horse is insanely happy

Junyang Lin (@JustinLin610) — Wed, 08 Apr 2026 04:22:19 GMT

happy horse is insanely happy

Chetaslua@chetaslua

🚨 Happy Horse First Output 

This model beats seedance 2 on artificial analysis for more information check quoted tweet

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬7🔄2❤️69👀29022📊13

⚡ Powered by xgo.ing

unbelievable...

Junyang Lin (@JustinLin610) — Wed, 08 Apr 2026 00:04:19 GMT

unbelievable...

Anthropic@AnthropicAI

Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
anthropic.com/glasswing

🔗 View Quoted Tweet

💬5🔄3❤️166👀38656📊21

⚡ Powered by xgo.ing

tokenmaxxing vs. ironmaxxing lol. it should be an era where results matter but it seems not.

Junyang Lin (@JustinLin610) — Tue, 07 Apr 2026 06:28:37 GMT

tokenmaxxing vs. ironmaxxing lol. it should be an era where results matter but it seems not.

💬9🔄2❤️74👀16923📊18

⚡ Powered by xgo.ing

Junyang Lin (@JustinLin610) — Sun, 05 Apr 2026 06:59:19 GMT

💬14🔄1❤️154👀17132📊30

⚡ Powered by xgo.ing

mountain climbing is so funny

Junyang Lin (@JustinLin610) — Fri, 03 Apr 2026 12:15:37 GMT

mountain climbing is so funny

💬13🔄1❤️107👀19500📊25

⚡ Powered by xgo.ing

model+harness is now over model only. agent perf can be significantly influenced by the design and q...

Junyang Lin (@JustinLin610) — Tue, 31 Mar 2026 00:59:46 GMT

model+harness is now over model only. agent perf can be significantly influenced by the design and quality of harness. i do believe this is a right direction, nice work!

Yoonho Lee@yoonholeee

How can we autonomously improve LLM harnesses on problems humans are actively working on?

Doing so requires solving a hard, long-horizon credit-assignment problem over all prior code, traces, and scores.

Announcing Meta-Harness: a method for optimizing harnesses end-to-end

🔗 View Quoted Tweet

💬22🔄53❤️573👀80722📊126

⚡ Powered by xgo.ing

https://t.co/mgVhhZVTKg

Junyang Lin (@JustinLin610) — Thu, 26 Mar 2026 10:35:54 GMT

x.com/i/article/2037…

💬88🔄587❤️3064👀895313📊1245

⚡ Powered by xgo.ing

nice work! better if oss...

Junyang Lin (@JustinLin610) — Mon, 23 Mar 2026 16:46:20 GMT

nice work! better if oss...

William Shen@shenbokui

UNI-1 is intelligent, directable, cultured. Incredible range it can do.

Incredibly proud of the world-class team building a world-class model.
It’s a daunting task to go up against industry giants like Deepmind/OpenAI/Bytedance.

More to come! API, technical report, model card…

Come join us!

🔗 View Quoted Tweet

💬15🔄9❤️423👀57797📊57

⚡ Powered by xgo.ing

by order of the peaky blinders

Junyang Lin (@JustinLin610) — Sun, 22 Mar 2026 10:02:05 GMT

by order of the peaky blinders

💬15🔄2❤️132👀23630📊30

⚡ Powered by xgo.ing

local agents will definitely become more important in the coming days and months while agents become...

Junyang Lin (@JustinLin610) — Sun, 22 Mar 2026 02:06:10 GMT

local agents will definitely become more important in the coming days and months while agents become part of our life and work. privacy always matters

Zach Mueller@TheZachMueller

PinchBench results for Qwen3.5 27B using @UnslothAI K_XL quants, best of 3, thinking enabled.

TL;DR: Q3 KXL (14.5GB) or Q4 KXL (18GB)

While overall the "best" results showed little degradation, if you dig into mean/std Q4_K_XL overall was the best at ~84% on average.

Q3 seems viable, while Q2 is the the lowest performing, of course.

🔗 View Quoted Tweet

💬35🔄38❤️568👀66084📊98

⚡ Powered by xgo.ing

working on distillation

Junyang Lin (@JustinLin610) — Sat, 21 Mar 2026 08:40:35 GMT

working on distillation

💬56🔄45❤️1644👀80912📊254

⚡ Powered by xgo.ing

this is a huge broccoli 🥦

Junyang Lin (@JustinLin610) — Thu, 19 Mar 2026 18:54:18 GMT

this is a huge broccoli 🥦

💬39🔄5❤️558👀42856📊101

⚡ Powered by xgo.ing

sry for missing messages. will respond asap

Junyang Lin (@JustinLin610) — Mon, 09 Mar 2026 05:17:34 GMT

sry for missing messages. will respond asap

💬97🔄11❤️822👀96966📊190

⚡ Powered by xgo.ing

me stepping down. bye my beloved qwen.

Junyang Lin (@JustinLin610) — Tue, 03 Mar 2026 16:11:24 GMT

me stepping down. bye my beloved qwen.

💬1691🔄713❤️13469👀6595655📊3756

⚡ Powered by xgo.ing

thx elon!

Junyang Lin (@JustinLin610) — Mon, 02 Mar 2026 19:19:38 GMT

thx elon!

Elon Musk@elonmusk

@Alibaba_Qwen Impressive intelligence density

🔗 View Quoted Tweet

💬33🔄39❤️1499👀380714📊173

⚡ Powered by xgo.ing

final shot. these are the small models that i told u!

Junyang Lin (@JustinLin610) — Mon, 02 Mar 2026 18:37:16 GMT

final shot. these are the small models that i told u!

Qwen@Alibaba_Qwen

🚀 Introducing the Qwen 3.5 Small Model Series
Qwen3.5-0.8B · Qwen3.5-2B · Qwen3.5-4B · Qwen3.5-9B

✨ More intelligence, less compute.
These small models are built on the same Qwen3.5 foundation — native multimodal, improved architecture, scaled RL:
• 0.8B / 2B → tiny, fast, great for edge device
• 4B → a surprisingly strong multimodal base for lightweight agents
• 9B → compact, but already closing the gap with much larger models
And yes — we’re also releasing the Base models as well.
We hope this better supports research, experimentation, and real-world industrial innovation.
Hugging Face:huggingface.co/collections/Qw…U
ModelScope:modelscope.cn/collections/Qw…I

🔗 View Quoted Tweet

💬44🔄41❤️1251👀149669📊161

⚡ Powered by xgo.ing

Junyang Lin (@JustinLin610) — Sun, 01 Mar 2026 11:39:34 GMT

💬66🔄38❤️1012👀169871📊203

⚡ Powered by xgo.ing

🤓

Junyang Lin (@JustinLin610) — Sat, 28 Feb 2026 00:25:47 GMT

🤓

Chen Cheng@cherry_cc12

Multimodal. Right-sized. It's a solid start for the 397B model, but not quite there yet. Keep improving, Chong!

🔗 View Quoted Tweet

💬5🔄0❤️65👀44354📊10

⚡ Powered by xgo.ing

OpenClaw + Qwen3.5

Junyang Lin (@JustinLin610) — Sat, 28 Feb 2026 00:03:04 GMT

OpenClaw + Qwen3.5

Alex Finn@AlexFinn

x.com/i/article/2027…

🔗 View Quoted Tweet

💬15🔄55❤️571👀120224📊172

⚡ Powered by xgo.ing

🧡ྀི💛ྀི❤️ྀི

Junyang Lin (@JustinLin610) — Fri, 27 Feb 2026 23:58:17 GMT

🧡ྀི💛ྀི❤️ྀི

Unsloth AI@UnslothAI

Qwen3.5 is now updated with improved tool-calling & coding performance!

Run Qwen3.5-35B-A3B on 22GB RAM.
See improvements via Claude Code, Codex.

We also benchmarked GGUFs & removed MXFP4 layers from 3 quants.

GGUFs: https://huggingface.co/unsloth/Qwen3.…s: https://t.co/IJuy70f4Vv

🔗 View Quoted Tweet

💬4🔄4❤️75👀28582📊12

⚡ Powered by xgo.ing

thx to nv for the support

Junyang Lin (@JustinLin610) — Fri, 27 Feb 2026 23:57:58 GMT

thx to nv for the support

NVIDIA AI Developer@NVIDIAAIDev

✨ Qwen3.5 — new from @Alibaba_Qwen — introduces a frontier‑class VLM built for native multimodal agents. 

With a ~400B‑parameter architecture combining MoE and Gated Delta Networks, Qwen3.5 can reason across text, code, and vision — and even understand and navigate user interfaces. 
 
Learn how to: 

✅ Run Qwen3.5 on free NVIDIA GPU endpoints 
✅ Deploy with NIM 
✅ Fine‑tune using NVIDIA NeMo 

See the details in our technical blog ➡️ developer.nvidia.com/blog/develop-n…

🔗 View Quoted Tweet

💬3🔄5❤️151👀25786📊18

⚡ Powered by xgo.ing

damn i love our medium size stuffs. things i use locally. oh try 122, sth new and sth quite nice!

Junyang Lin (@JustinLin610) — Tue, 24 Feb 2026 17:09:26 GMT

damn i love our medium size stuffs. things i use locally. oh try 122, sth new and sth quite nice!

Qwen@Alibaba_Qwen

🚀 Introducing the Qwen 3.5 Medium Model Series
Qwen3.5-Flash · Qwen3.5-35B-A3B · Qwen3.5-122B-A10B · Qwen3.5-27B

✨ More intelligence, less compute.
• Qwen3.5-35B-A3B now surpasses Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B — a reminder that better architecture, data quality, and RL can move intelligence forward, not just bigger parameter counts.

• Qwen3.5-122B-A10B and 27B continue narrowing the gap between medium-sized and frontier models — especially in more complex agent scenarios.

• Qwen3.5-Flash is the hosted production version aligned with 35B-A3B, featuring:
– 1M context length by default
– Official built-in tools

🔗 Hugging Facehuggingface.co/collections/Qw…jU

🔗 ModelScopmodelscope.cn/collections/Qw…CWI

🔗 Qwen3.5-Flash Amodelstudio.console.alibabacloud.com/ap-southeast-1…aqAF

Try in Qwen Chat 👇
Flchat.qwen.ai/?models=qwen3.…JZxIK

chat.qwen.ai/?models=qwen3.…4lETy

35B-chat.qwen.ai/?models=qwen3.…STbwh

122B-Achat.qwen.ai/?models=qwen3.…Xmh1o

Would love to hear what you build with it.

🔗 View Quoted Tweet

💬32🔄20❤️349👀22614📊63

⚡ Powered by xgo.ing

small coder in the coding plan now

Junyang Lin (@JustinLin610) — Fri, 20 Feb 2026 16:04:27 GMT

small coder in the coding plan now

Qwen@Alibaba_Qwen

Following the open-source release of Qwen3-Coder-Next, its API is now available on Alibaba Cloud Model Studio and has also been integrated into the Coding Plan.
For teams and developers who prefer scalable or cost-effective endpoints, you can now access Qwen3-Coder-Next via API.
🔗 API Documentation:modelstudio.console.alibabacloud.com/ap-southeast-1…D
🔗 Coding Plan Detailsalibabacloud.com/help/en/model-…i7
As always, feedback is welcome.

🔗 View Quoted Tweet

💬8🔄6❤️130👀17853📊20

⚡ Powered by xgo.ing

for some clarification i only mean fp8 amd it was released yesterday night

Junyang Lin (@JustinLin610) — Thu, 19 Feb 2026 12:08:35 GMT

for some clarification i only mean fp8 amd it was released yesterday night

Junyang Lin@JustinLin610

@QuixiAI will do soon

🔗 View Quoted Tweet

💬7🔄0❤️77👀10615📊12

⚡ Powered by xgo.ing

never an issue on qwen chat and no idea what happened with the deployment

Junyang Lin (@JustinLin610) — Thu, 19 Feb 2026 07:31:26 GMT

never an issue on qwen chat and no idea what happened with the deployment

Lucas@luksamuk

Eu me senti vendo o raciocínio do narrador de The Stanley Parable

flagship model sim, confia

🔗 View Quoted Tweet

💬5🔄0❤️59👀13155📊10

⚡ Powered by xgo.ing

bf16 too big? here it is the fp8 version that u been expecting for! try to enjoy the benefits of our...

Junyang Lin (@JustinLin610) — Wed, 18 Feb 2026 16:53:31 GMT

bf16 too big? here it is the fp8 version that u been expecting for! try to enjoy the benefits of our size and sparsity on ur device!

Qwen@Alibaba_Qwen

🚀 Qwen3.5-397B-A17B-FP8 weights are now open!
It took some time to adapt the inference frameworks, but here we are:
✅ SGLang support is merged
🔄 vLLM PR submitted github.com/vllm-project/v…Ti
Check the model card for example code. vLLM support landing in the next couple of days!
Hugging Facehuggingface.co/Qwen/Qwen3.5-3…ZR
ModelScopemodelscope.cn/models/Qwen/Qw…WK

🔗 View Quoted Tweet

💬14🔄7❤️268👀16841📊38

⚡ Powered by xgo.ing

we have our coding plan now and feel free to give it a shot

Junyang Lin (@JustinLin610) — Wed, 18 Feb 2026 15:00:53 GMT

we have our coding plan now and feel free to give it a shot

Qwen@Alibaba_Qwen

🚀 Qwen Coding Plan is now live on Alibaba Cloud Model Studio!

✨ What you get:
• 🔥 Latest Qwen3.5-Plus models
• 💡 Fixed monthly subscription: from ~$10/mo (Lite) or ~$50/mo (Pro)
• 📦 Up to 90K requests/month for AI-powered coding
• 🔌 Works with Claude Code, Qwen Code, Cline, Cursor & more
• 🔥 Get 50% OFF on Lite & Pro plans with our launch coupon!

🔗 Get staalibabacloud.com/help/en/model-…jcDUO7F

🔗 View Quoted Tweet

💬29🔄14❤️260👀21811📊52

⚡ Powered by xgo.ing

thx vercel for supporting qwen3.5

Junyang Lin (@JustinLin610) — Wed, 18 Feb 2026 07:52:40 GMT

thx vercel for supporting qwen3.5

Vercel Developers@vercel_dev

Qwen 3.5 Plus is now available on AI Gateway.

1 million token context window, built-in adaptive tool use, and strong visual reasoning.

Use 𝚖𝚘𝚍𝚎𝚕: '𝚊𝚕𝚒𝚋𝚊𝚋𝚊/𝚚𝚠𝚎𝚗𝟹.𝟻-𝚙𝚕𝚞𝚜'

🔗 View Quoted Tweet

💬5🔄1❤️72👀6531📊11

⚡ Powered by xgo.ing

@QuixiAI will do soon

Junyang Lin (@JustinLin610) — Wed, 18 Feb 2026 06:07:05 GMT

@QuixiAI will do soon

💬2🔄0❤️59👀15729📊10

⚡ Powered by xgo.ing

Qwen3.5 の日本語 OCR、気に入っていただけて嬉しいです！

Junyang Lin (@JustinLin610) — Wed, 18 Feb 2026 04:27:04 GMT

Qwen3.5 の日本語 OCR、気に入っていただけて嬉しいです！

ホーダチ-Hodatsu | LLM Researcher × AI Engineer@hokazuya

キタコレ
Qwen3.5-397B-A17Bの日本語OCR精度、これはやばいといわせてほしい

CER（文字誤り率）もひっくるめて、ほとんどの項目でGPT-5.2、Claude-Opus 4.6より高性能という結果に👀

OCRで正確な日本語抽出が難しいとされている、文部科学省のSSH支援事業の一枚を渡して同じ指示で出力させた結果、OCRの一般的な測定項目。
これがApache2.0で商用利用可能なモデルってのが激熱すぎる

もちろん、これだけで判定はできないけど、このレベルの画像に対する精度結果でこれだけの優位性はやばいでしょ

オンプレでB200購入して完全閉鎖空間でしかやれないPSの企業とちょうどOCRもローカルでという会話していたからありがたすぎる

テスト画像の出展：
文部科学省 - 令和6年度スーパーサイエンスハイスクール生徒研究発表会表彰校を決定しました - 参考1：スーパーサイエンスハイスクール概要（PDF:589KB）

🔗 View Quoted Tweet

💬8🔄23❤️366👀36290📊59

⚡ Powered by xgo.ing

cute!

Junyang Lin (@JustinLin610) — Tue, 17 Feb 2026 15:48:53 GMT

cute!

Chen Cheng@cherry_cc12

@test_tm7873 Qwen3.5

🔗 View Quoted Tweet

💬4🔄3❤️47👀4389📊8

⚡ Powered by xgo.ing

we recently have seen some unexpected cases in frontend and svg. we are still improving, but we use ...

Junyang Lin (@JustinLin610) — Tue, 17 Feb 2026 03:56:58 GMT

we recently have seen some unexpected cases in frontend and svg. we are still improving, but we use our system prompt and find that it works well in many cases. 

we have released our system prompt behind our web dev, and we welcome you to try out web dev in Qwen Chat.

Chen Cheng@cherry_cc12

Thank you and try in Qwen Chat and if you config OR API_KEY, it can work. It looks not bad.
Sys Prompt share:
You are a frontend developer. Output ONLY code, no explanations.

## Rules
 **Output**: Code block in markdown. No text before/after code.
 **Language**: Strictly match user's language
 **Format**: Single HTML code block with embedded CSS/JS, unless React/multi-file explicitly requested. If user requests React, output single file App.jsx.
 **Libraries**: Use CDN for external libs (p5.js, three.js, cannon-es, tailwindcss etc). For React projects, only use: recharts, framer-motion, react-three/fiber, react-three/drei.
 **Fonts**: Use CDN fontsource. Do not use google fonts.
 **Quality**: Bug-free, complete, runnable code. CRITICAL: Use unique variable names that don't conflict with library functions (e.g., avoid naming loop variable 'vertex' when using p5.js vertex() function).
 **Design**: Modern, clean UI with visual hierarchy, smooth animations, and responsive layout.
 **Completeness**: Implement ALL features requested. All buttons/interactions must be functional.
 **Images**: Use image generation tools to create image resources. Do not use unsplash or other online resources.
 **SVG**: For SVG, use SVG code block.

🔗 View Quoted Tweet

💬8🔄7❤️154👀15841📊29

⚡ Powered by xgo.ing

i really like this case

Junyang Lin (@JustinLin610) — Mon, 16 Feb 2026 11:13:18 GMT

i really like this case

Qwen@Alibaba_Qwen

Think, search, and create

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬5🔄3❤️122👀10214📊17

⚡ Powered by xgo.ing

A clarification of Qwen3.5 Plus and 397B: 1. for opensource, we follow the tradition to make parame...

Junyang Lin (@JustinLin610) — Mon, 16 Feb 2026 10:14:12 GMT

A clarification of Qwen3.5 Plus and 397B:

for opensource, we follow the tradition to make parameters apparent so we use the name with the number of total parameters and active params.

Qwen3-Plus is a hosted API version of 397B. As the model natively supports 256K tokens, Qwen3.5-Plus supports 1M token context length. Additionally it supports search and code interpreter, which you can use on Qwen Chat with Auto mode.

To help the opensource community, we also put the original 397B model for testing, which has thinking and fast (non-thinking) modes.

💬38🔄57❤️791👀57762📊164

⚡ Powered by xgo.ing

Qwen3.5 is Live! Today we openweight the first model, Qwen3-397B-A17B, which is a native multimodal ...

Junyang Lin (@JustinLin610) — Mon, 16 Feb 2026 09:43:41 GMT

Qwen3.5 is Live! Today we openweight the first model, Qwen3-397B-A17B, which is a native multimodal model supporting both thinking and non-thinking modes. We have strengthened its coding and agentic capabilities to foster productivity for developers and enterprises. Hope you enjoy it, and more open-weight models are coming in this Chinese New Year!

Qwen@Alibaba_Qwen

🚀 Qwen3.5-397B-A17B is here: The first open-weight model in the Qwen3.5 series.

🖼️Native multimodal. Trained for real-world agents.
✨Powered by hybrid linear attention + sparse MoE and large-scale RL environment scaling.
⚡8.6x–19.0x decoding throughput vs Qwen3-Max
🌍201 languages & dialects
📜Apache2.0 licensed

🔗Dive in:
Gitgithub.com/QwenLM/Qwen3.59joAT
Cchat.qwen.aiU0Rhwmodelstudio.console.alibabacloud.com/ap-southeast-1…TnHoU
Qwen Cgithub.com/QwenLM/qwen-co…nAger
Hugging Fhuggingface.co/collections/Qw…5p5um
ModelScmodelscope.cn/collections/Qw…Id57a
bqwen.ai/blog?id=qwen3.5StXaL

🔗 View Quoted Tweet

💬56🔄72❤️898👀60895📊150