Perplexity Decider v1.3 is out under Apache 2.0, and it never writes a sentence. It returns probabilities, and now it can also point at the exact words in your text that answer a question. It’s Monday 16 Nov, so this digest covers the 13 to 16 Nov news plus the weights that landed earlier in the week. I read the announcements, docs, model cards and pricing pages on 16 Nov. I ran none of it, so every number belongs to the source named next to it.
The useful stuff this week sits in the boring middle of the stack. A classifier, a search tool, a file parser.
🎯 What does Perplexity Decider v1.3 do if it never writes?
It answers questions about your content with numbers. Perplexity released Decider v1.3 and the smaller Decider Flash v1.3, distilled from it, around 9 to 10 Nov (the Hugging Face repo was created on 9 Nov). They succeed v1, v1.1, v1.2 and Flash v1.2.
Perplexity’s docs list four question types. noul gives a yes/no probability. choice gives a probability per option, up to 255. score gives a probability for each rubric level, up to 10, plus an expected score.
The new one is locate. It hands back the exact matching text from your input with start and end character offsets and a score per span. No match, empty list. Perplexity pitches it for named entity recognition, spotting personal information and extractive question answering. It is v1.3 only, text only, and does not work with web grounding.
Per the model card, Decider v1.3 is a 27B multimodal model on a Qwen3.8-27B base and needs about 49 GiB of BF16 weights plus working memory on a CUDA GPU. Flash is 4B. Both take 256K tokens of context.
The price is the headline for me. The docs list $0.02 per 1M input tokens for v1.3 and $0.01 for Flash, with output tokens free. Web grounding adds $1 per 1,000 searches or URL fetches. Same as v1.2. One catch from the changelog: pplx-decider-latest now points to v1.3, and probabilities can differ from v1.2. Tuned thresholds need a fresh check.
Perplexity’s own evaluations, from its launch post: across 63 classification tasks, with references relabeled independently by a frontier LLM, Decider v1.3 averages 91.6% accuracy, 6.54 points above Perplexity’s production classifiers. On 72 retrieval tasks it posts a mean nDCG@10 of 83.99, which is 1.81 points above Qwen3-Reranker-8B and 2.24 above Jev 1.14. Perplexity ran all of this itself.
That last name is where I have some history. The post says Decider is “inspired by Jev, TypeSafe AI’s System One Model.” I wrote a whole series on TypeSafe’s Jev and the System One idea, including a piece on how Noul, Choice and Score work. Now Perplexity ships the same primitives, down to the noul name. Probabilities instead of prose used to be a niche bet. It has a big company’s logo on it now.
Takeaway: at two cents per million input tokens, asking a model a yes/no question about every row in a table becomes a normal line item. Perplexity suggests v1.3 when mistakes are expensive and Flash for high-volume routing and labeling. Locate is the newest piece, so test the offsets on your own messy text first.
🔎 Can DeepSeek’s app features be an API call?
Since 13 Nov, yes. DeepSeek released “the full suite of application APIs”, so what its chat app does now works for deepseek-flash (DeepSeek-V4.1-Flash-1017) across the Messages, Responses and Chat Completions APIs, with tool calls and context assembly handled on DeepSeek’s servers.
Per the change log and the web tools guide, that covers web search with citations (deepseek_web_search, thinking off), a deep thinking search that reads pages across several reasoning steps (thinking on), URL reading via deepseek_read_url, file input for PDFs, Office files and images through POST /files and a file_id, and a safety check parameter. Search and Fetch, OCR, ASR and TTS also ship as standalone APIs.
The file parsing guide has the details you usually learn the hard way. The upload response only arrives after parsing, so raise client and proxy timeouts. You can never download the original or the parse result. Max 20 documents per request.
The pricing page gives deepseek-flash 1M context and 384K max output. Per 1M tokens, cache-miss input is $0.15 off-peak and $0.30 peak, output $0.60 and $1.20. Peak means 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Search and web fetch cost $0.75 per 1,000 calls each, and file parsing and OCR $0.225 per 1,000 pages.
Then the paragraph most people will skip. Public-facing apps must enable the safety checks, and developers stay responsible for content review, large model registration and AI labeling. When risk is detected, the API returns a refusal marker.
Takeaway: if you already run on deepseek-flash, you can delete your own search tool, fetcher and PDF parser. Searches bill per use on top of tokens, though, and an agent that searches every turn adds up fast. Read the compliance paragraph before shipping anything public. It is part of the product.

🎤 Why is a voice model that sings a tool-calling story?
Tencent Hunyuan released Hy Realtime, a real-time voice conversation model, on 13 Nov, as Yicai and IT之家 reported. The new part is agentic tool calling. The model decides from the conversation whether to run a web search, and the search decision, tool call and answer all happen inside one model.
Singing came out of work with Tencent Music’s TME Lingyin AI team. You can ask by singer, album, style, language, year or popularity. Tencent’s example: “an English folk song between 2010 and 2020”.
Tencent also claims better intent recovery from mumbled or self-corrected speech, from internal A/B and blind tests. No scores.
It is live in Yuanbao voice calls (the “Chuyu” voice) and Tencent customer service. A Tencent Cloud API is coming “soon”. No model id, parameter count, context window or price yet, per this roundup, and no weights.
Voice models are now sold the way text models already are, as agents that call tools. Until there is an API and a price, there is nothing to benchmark.

📊 Which of these can you actually download?
Prices are in US dollars per 1M tokens unless a row says otherwise.
| Model | Who | Date | Weights and license | Price or access | One number (source) |
|---|---|---|---|---|---|
| Decider v1.3 / Flash v1.3 | Perplexity | 9 to 10 Nov | Yes, Apache 2.0 | $0.02 / $0.01 per 1M input tokens, output free | 91.6% on 63 tasks (Perplexity) |
| deepseek-flash app tools | DeepSeek | 13 Nov | n/a (API feature) | search $0.75 per 1,000, tokens $0.15 / $0.60 off-peak | 1M context (DeepSeek docs) |
| Hy Realtime | Tencent | 13 Nov | No, closed | Yuanbao and customer service, API “soon” | none published |
| LongCat-2.5 | Meituan | weights 9 Nov | Yes, MIT | $0.30 / $1.20 limited-time | 48B active of 1.6T (Meituan) |
| Inkling 2 Small | Thinking Machines | weights about 12 Nov | Yes, Apache 2.0 | 3rd-party APIs | 17B active of 594B, 500GB+ VRAM (model card) |
| Seedream 5.5 | ByteDance | 9 Nov | No | Jimeng, Doubao, Ark, API pending | 0.4% severe text errors (ByteDance) |
📦 Did the big weights finally arrive?
Meituan put the LongCat-2.5 weights on Hugging Face under MIT on 9 Nov, a week after the launch, per VentureBeat. The Decoder reports that a separate 4-bit version and a self-hosting guide target a cluster of 194 Huawei Ascend chips. VentureBeat notes Meituan’s benchmark baselines are older than Opus 5.5 and GPT-6.1 Sol, so this is a sub-frontier model.
When I wrote about Inkling 2 Small, you could tune it but not download it. The weights went up on Hugging Face under Apache 2.0 around 12 Nov, per AI Compass. The card says the NVFP4 checkpoint needs at least 500GB of aggregated VRAM, for example 2x B300 or 4x H200, and lists NVIDIA, Fireworks, Together and others as API hosts.
“Open weights” and “you can run it” are two different sentences. 500GB of VRAM means a multi-GPU server at the very least, and 194 Ascend chips is a data center.
🖼️ What does a 100-round image edit mean?
ByteDance Seed released Seedream 5.5 on 9 Nov. The big change is multi-turn editing that keeps unedited regions at pixel level with no color shift, for up to 100 rounds on one image. It also outputs native RGBA with full and partial transparency, so no separate background removal step.
ByteDance reports a severe text corruption rate of 0.4% across Chinese and English tests and claims a lead over GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure. The Decoder and TNW both present these as ByteDance’s claims, with no independent check.
It is rolling out through Jimeng, Doubao, Doubao Work and Volcengine Ark, a Pro version is planned, and BytePlus says the API is “being integrated next”. Transparent PNGs and edits that don’t drift won’t trend. They do remove a step from real design work.
👀 What should you watch next?
Decider’s Locate on real, messy documents. DeepSeek search bills once agents call it every turn. A date and a price for Hy Realtime on Tencent Cloud. The Seedream 5.5 API and Pro. And whether anyone runs Inkling 2 Small or LongCat-2.5 on something smaller than a cluster. For the other cheap-model story, see last week’s digest with Meta’s ten cent worker model.
Sources
- https://www.perplexity.ai/hub/blog/perplexity-decider
- https://docs.perplexity.ai/docs/decisions/quickstart
- https://docs.perplexity.ai/docs/resources/changelog
- https://huggingface.co/perplexity-ai/pplx-decider-v1.3
- https://api-docs.deepseek.com/news/news261113/
- https://api-docs.deepseek.com/updates.md
- https://api-docs.deepseek.com/guides/web_tools/
- https://api-docs.deepseek.com/guides/document_parsing/
- https://api-docs.deepseek.com/quick_start/pricing
- https://www.yicai.com/news/103440081.html
- https://www.ithome.com/1/026/825.htm
- https://aiwebcool.com/en/posts/hy-realtime-voice-model-hotspot
- https://venturebeat.com/technology/meituan-releases-the-open-weights-of-its-longcat-2-5-llm-with-vision-under-a-permissive-mit-license
- https://huggingface.co/meituan-longcat/LongCat-2.5
- https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/
- https://ashutoshai.com/ai-compass/2026/11/12/thinking-machines-lab-inkling-2-small-open-weights/
- https://huggingface.co/thinkingmachines/Inkling-2-Small
- https://seed.bytedance.com/en/blog/more-realistic-images-more-stable-editing-across-100-rounds-introducing-seedream-5-5
- https://seed.bytedance.com/en/seedream5_5
- https://the-decoder.com/bytedances-seedream-5-5-promises-pixel-stable-image-editing-across-100-rounds-without-color-drift/
- https://thenextweb.com/news/bytedance-seedream-5-5-image-editing-100-rounds
- https://ai.byteplus.com/lumina/en/resource/seedream-5-5-release-date