🚨 DeepSeek V4-Flash 0731, MiniMax H3, Supabase evals

Presented by MongoDB  Event Venue Operator — an agentic demo on Atlas with three‑layer memory and Voyage embeddings. Clone it →

Almost everything below left the model alone. DeepSeek shipped the same 284B architecture with new post‑training. Tencent froze the target and trained a drafter around it. Liquid AI converted decoders it already had. Supabase changed the skills, not the agents. The exception is MiniMax H3, which rebuilt the whole stack — and is the one thing here you cannot download.

82.7

up from 61.8 preview

4.79

against 3.00 for MTP

99.4%

tokens saved, 233 pages

01 · DeepSeek · Model update

DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains

An upgrade usually means a new architecture. This one means a new curriculum. The model card says the architecture and size are unchanged. The gains come from re‑post‑training alone.

Terminal Bench 2.1 moves 61.8 to 82.7. DeepSWE moves 7.3 to 54.4. It beats V4‑Pro (Preview) on every agentic benchmark DeepSeek published. Output pricing is $0.28 per 1M tokens against $0.87 for deepseek-v4-pro. Weights are MIT and ungated, but self‑hosting still wants roughly 110 GB at 3‑bit. Every figure is vendor‑reported on a harness that has not been released.

Read on Marktechpost →  · Model card  · Technical report  · Pricing  · DSpark paper

02 · MiniMax · Video model

MiniMax Releases MiniMax H3: An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio

A video stack used to be a drawer of expert models. Text‑to‑video, image‑to‑video, first‑and‑last‑frame, subject reference, editing. H3 folds them into one pretraining paradigm where the relationships are expressed in natural language.

Output is 2K, four to fifteen seconds, integer durations only, with native stereo. H3‑VAE’s compression delivers a stated 4× gain in effective sequence length, which is what makes native 2K economical. Training throughput is up nearly 30%. Open weights are promised in the coming days, not shipped, so the API is the only path today.

Read on Marktechpost →  · Announcement  · API docs

03 · Supabase · Benchmark

Supabase Releases Evals: an Open Source Benchmark That Scores Claude Code, Codex and OpenCode on Real Supabase Tasks

Most agent benchmarks grade against mocks. This one boots a containerized stack and a real CLI, then asks the agent to build a schema or fix a broken RLS policy. Scoring mixes deterministic checks with LLM‑as‑a‑judge, and the agent gets one retry.

Agents pass most scenarios with no skill loaded. Opus 5 and Kimi K3 both scored 100% unaided in the Build stage. Skills mattered further down: Sonnet 5 went 78% to 100%, GPT‑5.4 mini 78% to 89%. The odd finding is docs usage. Codex reads roughly eight docs pages per scenario against about two for Claude Code. Apache‑2.0, and it runs locally on pnpm.

Read on Marktechpost →  · Announcement  · GitHub  · Leaderboard

From our sponsor
Want agent memory modelled on a real cluster instead of a diagram? Run MongoDB’s Event Venue Operator →

04 · Tencent · Inference

Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models

Speculative decoding usually hunts for one drafter that averages well. Real traffic is not an average. Open‑ended chat rewards short proposals. Code and maths reward long predictable blocks. AngelSpec ships two drafters instead of one compromise.

DFly reaches 4.79 mean accepted length on Hy3‑A21B, against 3.69 for DFlash and 3.00 for MTP. On Qwen3‑8B it takes every math and code benchmark, though DSpark stays ahead on MT‑Bench at 3.77 to 3.67. The target backbone and LM head stay frozen throughout, so verification preserves the original distribution. Seven checkpoints ship on Hugging Face and ModelScope.

Read on Marktechpost →  · Paper  · GitHub  · Docs  · Hugging Face

05 · Liquid AI · Encoders

Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU

Nobody demos an encoder. Everybody runs one, underneath the classifier, the intent router, the PII filter. Liquid AI did not train these from scratch. It converted decoders it already had, using a bidirectional mask, non‑causal short convolutions, and a 30% mask rate.

The 350M posts an 81.02 mean across 17 tasks, ranking fourth of 14. The three models above it are all larger, and the top one is nearly ten times its size. The 230M runs one 8K forward pass on CPU in about 28 seconds, against over 90 for ModernBERT‑base. License is the LFM Open License v1.0, and every load call needs trust_remote_code=True.

Read on Marktechpost →  · Announcement  · Model card  · Eval code

06 · Marktechpost · Open source

Meet Token Saver: An Open-Source MCP Extension Using Local Hybrid RAG to Cut Claude PDF Token Costs

A 200‑page PDF is not a one‑time charge. Conversation history is re‑sent on every turn, so the document is paid for again with every follow‑up question. Token Saver keeps the file on your machine and sends only the passages that answer the question.

BM25 over SQLite FTS5 at weight 0.4, local all‑MiniLM‑L6‑v2 cosine at 0.6, 180‑word chunks, and an 8,000‑character cap on what reaches the model. A 233‑page court opinion goes from 133,349 tokens to 740, a 99.4% saving. GDPR at 88 pages saves 98.6%. MIT licensed, one .mcpb bundle, no Python environment.

Read on Marktechpost →  · GitHub  · MCP Bundle format  · Embedding model

One more thing…

Sponsored · MongoDB · Agentic AI

Event Venue Operator: an agentic demo you can clone, built on MongoDB Atlas

Most agent demos are a screenshot. This one is a repo. An AI operator runs day six of a tennis tournament with rain ninety minutes out, and has to re‑plan for two guests who want opposite things.

Three memory layers sit on a single Atlas cluster. Long‑term lives in memory_store with Voyage embeddings, short‑term holds venue and weather state, and a shared agent_actions log coordinates subagents. Retrieval is Atlas Vector Search plus hybrid lexical search.

The agent is a real LangGraph chain — perceive, plan, hitl_gate, act, reflect — not a scripted walkthrough. Embeddings are voyage-multimodal-3.5 at 1024 dimensions, cross‑modal over text and images. Langfuse tracing is optional. MIT licensed, Python 3.12+, and there is a hosted demo if you want to look before you clone.

Repo & README → · Live demo · Full tutorial

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See a placement in action — this issue’s feature for MongoDB → That is what your slot looks like. Ask for rates and open dates →