DeepSeek 멀티모달 API, Qwen 추론 레시피, RULER 학습 신호와 장편 영상·코드 에이전트 구현

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8. Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model. 1/n

Multimodal API support Set model='deepseek-v4-flash-vision-exp' Images are tokenized for billing: up to 384 tokens each, at V4-Flash pricing Supports Chat Completions, Messages & Responses Supports mixed text + image input; images can be provided via base64, external URLs, or the Files API. Docs: https:// api-docs.deepseek.com/guides/vision 3/n

Fresh recipes just dropped! NVFP4 + DFlash2 for Qwen3.8-27B now in the SGLang cookbook. Thanks for the support! @sgl_project

I have been fine-tuning LLMs for over 2 years now! Here are the top 12 techniques I'd learn if I were to fine-tune them: Bookmark this. 1. LoRA > Freezes the base weights and trains two low-rank matrices as the update, resulting in ~95-99% fewer params to fine-tune. 2. QLoRA > LoRA on top of a 4-bit quantized base model. 3. Prefix tuning > Prepends trainable vectors to keys and values at every layer, weights frozen. 4. Adapter tuning > Inserts small trainable modules between transformer layers. 5. Instruction tuning > Supervised tuning on (instruction, response) pairs so the model follows directions instead of just continuing text. 6. P-tuning > Optimizes continuous prompt embeddings through a small encoder, mainly for NLU tasks where discrete prompts are unstable. 7. BitFit > Trains only the bias terms, ~0.08% of params, and still rivals full fine-tuning on small-to-medium datasets. 8. RLHF/RLAIF > RLHF trains a reward model on human preference rankings, then PPO against it. The pipeline behind the first ChatGPT. > RLAIF swaps the human labeler for an LLM judging. RLHF-level quality at a fraction of the cost. 9. DPO (Direct Preference Optimization) > Skips the reward model and optimizes preference pairs directly with a classification-style loss. Simpler than PPO. 10. GRPO (Group Relative Policy Optimization) > Samples a group of responses per prompt and normalizes their rewards within the group. DeepSeek R1 ran on it. 11. RLVR (Reinforcement Learning with Verifiable Rewards) > Replaces the learned reward model with a checker or compiler returning verifiable scores. The free signal behind R1's math and code. 12. Federated fine-tuning > Tunes across decentralized clients that share only weight updates, never raw data. For when data can't leave the device. GRPO needs exactly one scalar reward per response. RLVR (13) produces that for free on math and code by running the answer through a checker or compiler. But tasks like a RAG answer, a support reply, or a summary have no gold label to match against. The usual fallback is a hand-written reward function scoring faithfulness, hallucination, and completeness. It takes days to calibrate, rewards the wrong behavior when the weights are off, and breaks every time you add a tool or edit the system prompt. RULER, implemented in OpenPipe's ART (open-source), solves this. During training, it passes the N sampled trajectories to a judge LLM, which ranks them relative to each other against the agent's system prompt and returns the scores. Relative ranking is more stable than absolute scoring, and GRPO normalizes within the group anyway, so the rankings feed straight into the pipeline like with RLVR. Here's the GitHub Repo: https:// github.com/OpenPipe/ART (don't forget to star it ) I wrote a full breakdown recently on how exactly this works, with the training loop and code. Read it below.

Introducing HyCreator, our agent harness for long video generation. Most video models hand you a clip and wish you luck. This one runs the whole thing end to end, and lets you cut in at any point instead of re-rolling the prompt

Claude Code is now scary good at full-stack! I asked it to build a real-time weather intelligence dashboard with an interactive 3D globe, a forecasting layer that predicts weather 3 days ahead, and an anomaly detector that flags cities whose weather is behaving abnormally. It came back with a spinning globe that has a day/night cycle using NASA satellite imagery, city lights on the dark side, weather icons that switch between sun and moon based on local time, and a time travel slider that scrubs through 10 days of data. And when a city's weather breaks from its own normal, it pulses red (abnormally hot) or blue (abnormally cold) right on the globe, updating live and reflecting the anomaly state at any point you drag the slider to. Claude Code built the whole thing in a single session, including the backend, database, data pipeline, and frontend. For the database, I needed something fast for time-series workloads since the app ingests hourly weather readings across many cities and serves time-range queries on every slider interaction. I used Tiger Cloud by @TigerDatabase , which gives you managed TimescaleDB on the Postgres you already know. Claude Code connected to it through the Tiger CLI MCP server and set up the entire backend directly: - Provisioned the database service - Created hypertables for time-partitioned weather storage - Set up continuous aggregates for pre-computed rollups - Built the data ingestion pipeline and the full NextJS + ThreeJS frontend The time travel slider queries thousands of rows on every position change. On a regular Postgres table, this would require manual partitioning and index tuning to stay fast as data grows. TimescaleDB partitions the data by timestamp automatically, so each query only hits the relevant time chunk. Continuous aggregates serve the trend charts, the forecast layer, and the anomaly baselines from pre-computed rollups instead of rescanning raw data on every request. The video below shows the final build in action, and I worked with the Tiger Data team to put this together. Tiger CLI is open-source (Apache 2.0) and works with Claude Code, Cursor, Codex, Gemini CLI, and VS Code. To try this yourself: → Sign up for Tiger Cloud (I have shared the link in the replies). It gives you $1,000 free credits (no card needed) → Install Tiger CLI: curl -fsSL https(:)//cli(.)tigerdata(.)com | sh → Run tiger mcp install claude-code → Give Claude Code a prompt and let it build sign-up here: https:// fandf.co/4gFoRHn My co-founder also wrote a detailed article on this. The article is quoted below.

jietang

@jietang

14일 전

cool comment: FLOPs were intelligence; parameters were knowledge!