Skip to content

#开源/仓库

3 today

Apr 20Monday

Hacker News front page

Show HN: TRELLIS.2 image-to-3D running on Apple Silicon, no Nvidia GPU needed

Developer shivampkumar ported Microsoft's 4B-parameter TRELLIS.2 to Apple Silicon with PyTorch MPS for single-image 3D generation. He replaced flash_attn, nvdiffrast, and custom sparse conv kernels with pure PyTorch sparse 3D conv, SDPA attention, and Python mesh extraction. On an M4 Pro with 24GB, it generates ~400K-vertex meshes in about 3.5 minutes; slower than H100 seconds, but fully offline.

Why it matters: Strong on all HKR axes: a clear hook, concrete implementation details, and benchmark-like numbers. This is not a Microsoft model launch, but a reproducible local port with real practitioner relevance, so it lands in featured rather than p1.

Apr 17Friday

Xinzhiyuan · WeChat

Yixin says its finance Agent harness runs single tasks for 16 hours and plans an H2 open-source release

Yixin says its finance Agent harness can run a single task for 16 hours across 12 sessions, with 65% autonomous delivery. The post adds a 50k-token cap per case, projected approval speedups above 150%, and projected unit cost at one-fifth of human work; it says an open-source release is planned for H2 2026, but does not disclose the repo, license, or reproducible evals. The key signal is governance design, not the “smarter over time” framing.

Why it matters: This clears HKR-H/K/R with a rare production claim: a finance agent runs 16 hours, spans 12 sessions, hits 65% autonomous delivery, and stays under a 50k-token cap. It stays below 85 because the evidence is self-reported and the post does not disclose a repo, license, or reproduc

Hacker News front page

Discourse Is Not Going Closed Source

Discourse said it will keep its GPLv2 codebase open after 13 years. The post says its team used GPT-5.3 Codex, GPT-5.4, and Claude Opus 4.6 to scan code, and its last monthly release fixed 50 security issues. The key claim is defensive capacity: OpenAI said Codex Security scanned 1.2M+ commits in 30 days and found 792 critical and 10,561 high-severity issues.

X · @dotey

browser-use open-sources video-use, a Claude Code skill that turns raw camera footage into edited videos

browser-use released video-use, a Claude Code skill that turns raw footage into a final.mp4 automatically. It converts footage into ElevenLabs word-level timestamp transcripts, shrinking one asset to about 12KB; the post says feeding frames directly would cost about 45 million tokens. The key detail is the structured editing pipeline: the model mostly reads text, uses timeline images only at uncertain cuts, and runs up to 3 self-check repair passes after rendering.

Why it matters: Strong HKR-H/K/R: the result is instantly clickable, and the post includes a concrete text-first editing architecture with 12KB vs about 45M-token economics. Kept below higher bands because this is a builder-facing Claude Code skill, not a platform-level release.

Apr 15Wednesday

X · @dotey

pi maintainer Mario Zechner sets a new rule: unapproved issues and PRs will be auto-closed immediately

pi maintainer Mario Zechner says any issue or PR submitted without prior approval will be auto-closed, after he started receiving 30 to 50 issues per day and most were AI-agent spam. He will still review closed submissions daily; strong issues can earn an “lgtmi” tag, and strong issue-plus-fix PRs can earn “lgtm,” exempting future submissions from auto-close. The shift to watch is simple: open source projects are raising contribution gates to filter zero-cost AI-generated noise.

Why it matters: Featured on strong HKR-H/K/R: a maintainer-level policy change with concrete spam numbers and a review mechanism. Importance stays in the mid-70s because the blast radius is mainly the OSS agent/dev community, not a major model or platform release.

Apr 14Tuesday

X · @dotey

Vercel open-sources Open Agents, a reference implementation for enterprise coding agent platforms

Vercel open-sourced Open Agents as a forkable reference for enterprise coding-agent platforms, with a three-layer architecture and features like voice input and PR creation. Its key design keeps the agent outside the sandbox and uses tools such as file I/O, shell, and search to control execution; the post also cites Anthropic Managed Agents pricing at $0.08 runtime per hour and $10 per 1,000 web searches. The part to watch is the agent-sandbox split, not the packaging choice.

Why it matters: This fits the 78–84 band: a notable open-source coding-agent framework with concrete architecture, remote sandbox operation, and Anthropic pricing, so HKR-H/K/R all land. It stops short of must-write status because this is strong infra reference material, not a model or industry-

Apr 12Sunday

X · @Yuchenj_UW

MiniMax M2.7 is open-source!

MiniMax open-sourced M2.7 and said its research agent now handles 30%–50% of the R&D workflow. The post says the agent covers literature review, experiment orchestration, log debugging, code fixes, and merge requests; M2.7 also rewrote its own harness for 100+ automated rounds, with a 30% gain on internal coding evals.

Why it matters: HKR-H/K/R all pass: open-sourcing plus a research agent doing 30%-50% of R&D is a strong hook, and the post includes 100+ self-rewrite loops with +30% internal coding eval. It stays at 78 because license, repo, benchmark context, and external reproduction are not disclosed.

Apr 11Saturday

QbitAI · WeChat

Liu Zhuang and Danqi Chen team open-source Vero, a general visual reasoning RL framework, reaching SOTA with zero thinking data

Princeton researchers including Liu Zhuang and Danqi Chen open-sourced Vero, an RL framework for visual reasoning, and report beating Qwen3-VL-8B-Thinking on 23 of 30 benchmarks. The post says Vero uses 600K samples filtered from 59 datasets, task-routed rewards, and single-stage RL across six task groups. The key point is the mechanism mix: no private thinking data, but the post does not disclose training cost or base model configuration.

Why it matters: Featured on HKR-H/K/R: the zero-thinking-data claim is a strong hook, and the post includes concrete benchmark and method details. I keep it in the low 80s because training cost, base model choice, and full reproduction conditions are not disclosed.

Apr 10Friday

最佳拍档 (BestPartners)

LLM self-evolution: Shinka Evolve, AlphaEvolve, and sample efficiency

Sakana AI open-sourced Shinka Evolve and uses a UCB bandit to switch among GPT-5, Claude Sonnet 4.5, Gemini, and others, aiming to cut the thousands of program evaluations common in AlphaEvolve-style search. The post says it beat AlphaEvolve’s classic circle-packing result with fewer evaluations and adds full-file rewrites, crossover, editable-region guards, and a meta-notebook; the post does not disclose exact metrics, cost, or the repo link. The part to watch is surrogate-task design and hard verification: the system still needs humans to define problems.

Why it matters: Featured, not P1: HKR-H/K/R all pass. The piece has a strong hook, concrete mechanisms like UCB model routing and program crossover, and a real nerve around eval cost and hard verification. It stays at 80 because key metrics, cost, and the primary release link are not disclosed.

QbitAI · WeChat

Tencent open-sources 3B SVG model HiVG to make tokens geometry-aware

Tencent Hunyuan open-sourced the 3B-parameter HiVG, claiming 62.7%-63.8% shorter SVG sequences via hierarchical tokenization and better SVG generation metrics than GPT-5.2, Claude-4.5-Sonnet, and some 8B open models. The post reports 0.896 SSIM, 0.114 LPIPS, and 0.957 CLIP-S on Image-to-SVG; the core method packs drawing commands plus coordinates into segment tokens and uses HMN to initialize coordinate embeddings. The part to watch is token design, not parameter count; paper, code, and project page are public.

Why it matters: Tencent's HiVG earns HKR-H and HKR-K: a 3B open model claims GPT/Claude-level SVG results, and the article includes 62.7%-63.8% token compression plus SSIM 0.896, LPIPS 0.114, and CLIP-S 0.957. HKR-R is weaker because SVG generation remains niche, so it lands at the low end of `f

Apr 8Wednesday

QbitAI · WeChat

Free open-source 2B Chinese speech model reproduces Mangzhuang Ren with high-speed tonguetwisters

ModelBest, OpenBMB, and Tsinghua University released VoxCPM 2, a 2B open speech model that supports 9 Chinese dialects, 30 foreign languages, and 48kHz audio. The post says generation often finishes within 1 second, recommends reference audio of at least 5 seconds, and supports denoising, LoRA, and full fine-tuning; the key detail is its tokenizer-free diffusion autoregressive continuous representation design.

Why it matters: This is a substantive open-source speech release, not a thin demo: the post gives 2B, 48kHz, 9 Chinese dialects, 30 languages, ref audio ≥5s, and a tokenizer-free route. HKR-H/K/R all pass, but the event is not large enough for a must-write P1.

X · @dotey

Hermes Agent is gaining traction; I installed it and the experience was decent

Nous Research open-sourced Hermes Agent in late February, and the post says it reached nearly 30,000 GitHub stars in under two months. The post describes a closed learning loop: after complex tasks with 5+ tool calls, Hermes writes Markdown skills, with one Reddit report claiming 3 skills in 2 hours and a 40% speedup on repeated research work. The key angle is its self-hosted agent engine that combines skill generation, SQLite-based memory retrieval, and five-layer safety controls.

Why it matters: HKR-H/K/R all pass: the piece combines strong OSS momentum, concrete mechanics, and a real builder nerve around self-hosted learning agents. It stays at 78 because the evidence is mostly social commentary and light user feedback, not a primary release or broad independent eval.

Apr 7Tuesday

X · @dotey

Milla Jovovich and Ben Sigman release open-source AI memory system MemPalace, claim perfect LongMemEval score

Milla Jovovich and Ben Sigman released the open-source memory system MemPalace and claimed a perfect LongMemEval score. The project runs fully local with no cloud or API key, says AAAK compresses context 30x, and uses 19 MCP tools for retrieval. The key issue is evaluation: Penfield Labs says the “perfect” result measured retrieval only, not end-to-end QA, and AAAK dropped retrieval accuracy from 96.6% to 84.2%.

Why it matters: HKR-H lands on the celebrity/open-source hook and the 'perfect score' dispute. HKR-K/R land on concrete metrics and the familiar nerve of eval gaming vs real memory utility; source authority is still just an X post, so this stays featured, not higher.

Latent Space

[AINews] Gemma 4 crosses 2 million downloads

Google’s Gemma 4 reached about 2 million downloads in its first week. The post compares that with Gemma 3 at 6.7 million over the past year, Gemma 2 at 1.4 million since June 2024, and Qwen 3.5 at about 27 million in roughly 1.5 months. The signal for practitioners is local deployment: one iPhone 17 Pro demo ran Gemma 4 E2B at about 40 tok/s via MLX, with support across Hugging Face, vLLM, llama.cpp, Ollama, and NVIDIA.

Why it matters: HKR-H/K/R all pass: the story has a clean hook, concrete comparative download data, and a real open-model adoption nerve. It stays low-featured because this is a secondary-source uptake snapshot, not a primary Google release or a substantive capability update.

Apr 3Friday

X · @dotey

Google releases the Gemma 4 open model family under Apache 2.0

Google released the Gemma 4 family and switched the full line to Apache 2.0. The post says it includes 31B Dense, 26B MoE, E4B, and E2B; 31B and 26B support 256K context, and 31B fits on one 80GB H100. The key change is distribution terms: fewer limits on commercial use, modification, and redistribution, plus native function calling and structured JSON for agent workflows.

Why it matters: This is a substantive Google model release, with the Apache 2.0 switch carrying as much weight as the model specs. HKR-H/K/R all pass on novelty, concrete deploy details, and commercial relevance; it stays below P1 because the post lacks formal eval links and direct head-to-heads

Feb 12Thursday

MIT Technology Review · AI

What’s next for Chinese open-source AI

MIT Technology Review says that after DeepSeek released R1 in January 2025, Chinese firms kept shipping open-weight models near top Western systems; Moonshot AI’s Kimi K2.5 was close to Anthropic Claude Opus on early benchmarks at about one-seventh the price. The post also says Qwen took over 30% of Hugging Face downloads in 2024 and surpassed Meta Llama in cumulative downloads by 2025–2026; the key shift is from a few general models to many fine-tunable, distillable variants.

Why it matters: All three HKR axes pass. This is not a launch, but it offers concrete market signals—~1/7 pricing, Hugging Face download share, and a clear thesis that Chinese open source is moving toward specialized, distillable variants—so it merits featured, not p1.

Lex Fridman (YouTube RSS)

OpenClaw: The Viral AI Agent Behind the Hype - Peter Steinberger | Lex Fridman Podcast #491

Lex Fridman’s episode #491 interviews Peter Steinberger about the open-source AI agent OpenClaw; the transcript says it reached 175k-180k GitHub stars. The post says it can connect to Telegram, WhatsApp, Signal, and iMessage, and use models such as Claude Opus 4.6 and GPT 5.3 Codex; it does not fully disclose the architecture, evals, or security boundaries. The real point is system-level access and self-modifying behavior: this is not chat, but an agent that can take actions.

Why it matters: This is more than a routine podcast. OpenClaw scores on HKR-H/K/R with 175k-180k GitHub stars, messaging integrations, and self-modifying behavior. It stays at featured, not p1, because the post does not disclose architecture, evaluations, or safety boundaries.

Aug 5, 2025Tuesday

OpenAI News

gpt-oss-120b & gpt-oss-20b Model Card

OpenAI released gpt-oss-120b and gpt-oss-20b as open-weight reasoning models under Apache 2.0, with compatibility for the Responses API. They are text-only models with tool use, Structured Outputs, and adjustable reasoning effort; the post does not disclose context length, pricing, or benchmark scores. On safety, OpenAI says gpt-oss-120b stayed below the High threshold in bio, cyber, and AI self-improvement tests, including after adversarial fine-tuning.

Why it matters: This is a same-day write: HKR-H from OpenAI going open-weight, HKR-K from license/mechanism/safety specifics, and HKR-R from the open-vs-closed debate. I kept it below 90 because the post excerpt does not disclose context length, pricing, or full benchmark results.