Choosing the right AI model is becoming more important than learning prompts. **General Productivity** • ChatGPT • Claude • Grok • Pi **Advanced Reasoning** • Claude • Gemini • DeepSeek • Mistral **Coding** • GitHub Copilot • Claude • ChatGPT • StarCoder2 **Creative Writing** • Claude • ChatGPT • Gemini • Mistral **Research & Search** • Perplexity • ChatGPT • Grok • You.com **Long Documents & Analysis** • Claude • ChatGPT • Gemini • DeepSeek **Multimodal Tasks** • ChatGPT • Gemini • Claude • Grok **Open-Source Models** • DeepSeek • Mistral • StarCoder2 • GPT4All **Education & Learning** • ChatGPT • Claude • Perplexity • DeepSeek The biggest mistake most people make is using one AI for everything. Match the model to the task, and you'll get better results, save time, and avoid unnecessary frustration. Which AI model is your daily go-to?

noteで記事を書きました！この投稿をリポストするとお得に記事を読むことができます。 starcoder2:7b Phase 2 日本語性能レポート | 電霞リリ & トウマ @LILI_DENKA #note note.com/la_e/n/n725ef69b203…

新しいnote記事を公開しました。「starcoder2:7b Phase 3 コーディング性能レポート」記事は150円で読めます。 note.com/la_e/n/n22881534653… #電霞リリ #生成AI #LLM #コード生成

Speaker: Noumane Tazi (@Nouamanetazi) • ML Engineer @ Hugging Face (@huggingface) • Lead author, Ultra-Scale Playbook • Core dev of Nanotron • Worked on StarCoder2, SmolLM3, MoE scaling (4/6)

Speaker Spotlight at #GOSIMParis 2026 🎤 👨Nouamane Tazi Machine Learning Engineer @ Hugging Face Topic: Optimizing & Scaling Open-Source LLMs 💡What you’ll learn: • GPU optimization & 3D-parallelism with Nanotron • Building & training efficient code LLMs (StarCoder2) • Deploying SOTA open models in production • Practical techniques for low-resource LLM training 📅GOSIM Paris 2026 • May 5–6 • Station F 👉Hear Nouamane Tazi speak at GOSIM Paris for just €10.9 — see how? pinned post ↑ #GOSIMParis #OpenSource #DeveloperCommunity #CloudNative #Kubernetes #Rust #AI #WebAssembly #DevRel #ParisTech #TechConference #OpenSourceCommunity @Huawei @huggingface @AIFC_Official_ @Kimi_Moonshot @Minimax @Openharmony @OpenBMB @Zhipu_AI @Nouamanetazi

[論文]GitHub CopilotやCursorなどAIコーディングアシスタントの背後にあるコード生成LLMが、訓練データの汚染によって脆弱なコードを出力するよう仕込まれていた場合に、モデルの内部にアクセスせずに外側から検出するフレームワーク「CodeScan」の提案。 汚染されたモデルは、たとえばFlaskアプリのテンプレート描画を安全なrender_template()ではなくXSSに脆弱なjinja2[.]Template()[.]render()で書くよう誘導されるといった形で、一見正常だが危険なコードを生成する。 従来のスキャン手法BAITはトークン単位の一致を見るため、変数名やファイル名が変わるだけで同じ脆弱な構造を見落とす一方、無害な定型コードを誤検知してしまう問題があった。 CodeScanは生成コードを抽象構文木(AST)に変換して表層的な違いを吸収し、複数のプロンプトにまたがって繰り返し出現する構造パターンを抽出したうえで、そのパターンに脆弱性が含まれるかをLLMで判定する仕組み。 【要点】 ・University of ConnecticutとVisa Researchによるプレプリント(arXiv:2603.17174、2026年3月17日公開)。コード生成LLMを対象としたブラックボックスかつ脆弱性指向のポイズニングスキャンとしては初の提案とされる ・CodeLlama、Qwen2.5-Coder、StarCoder2の3アーキテクチャ、108モデルに対し4種の攻撃手法(Simple、Covert、TrojanPuzzle、CodeBreaker)で評価。検出F1スコアは平均約0.98 ・比較対象のBAITはコード生成LLMに適用するとV1(XSS)およびV2(証明書検証無効化)で偽陽性率100%、全体のF1スコアは0.17。CodeScanはV1で偽陽性率0%、V2・V3で16.67%に抑制 ・CodeBreaker攻撃はGPT-4ベースの検出器を回避するよう難読化されたペイロードを使うが、CodeScanはGPT-5 miniを脆弱性判定に使うことでこの回避も突破。GPT-5 miniは難読化済みコードに対してもzero-shotで94%以上、one-shotでほぼ100%の検出精度を発揮 ・スキャン時間もBAITと比較して大幅に短縮されており、実用面でも優位 ・34Bを含む大型モデルでも高い精度・再現率を維持しており、モデル規模へのスケーラビリティも確認済み arxiv.org/abs/2603.17174

Open Source AI Models and Their Specialties: Large Frontier Models > Meta Llama 4 (Maverick, Scout) — multimodal, 10M token context window > DeepSeek V3 / R1 — rivaling closed models at a fraction of the cost > Mistral Large / Mixtral — efficient mixture-of-experts architecture > Qwen 2.5 (by Alibaba) — strong multilingual and math performance > Google Gemma 2 — lightweight but competitive with much larger models Coding-Focused > DeepSeek Coder V2 — top open-source coding benchmark scores > Qwen2.5-Coder — strong across code generation and debugging > StarCoder2 (BigCode) — trained on 600 programming languages > CodeLlama — Meta's code-specialized Llama variant Small/Efficient Models > Meta Llama 3.2 (1B, 3B) — runs on-device, mobile-ready > Microsoft Phi-3 / Phi-4 — punches way above its weight class > Google Gemma 2 (2B, 9B) — great for local deployment > Mistral Small — fast inference, low resource requirements Reasoning Specialists > DeepSeek R1 — chain-of-thought reasoning, open-weight > Qwen QwQ — dedicated reasoning model with transparent thinking

Early Statistical & Neural LMs (pre-2018) n-gram models (various, 1990s–2000s) — Kneser-Ney, Good-Turing smoothed models Word2Vec (Google, 2013) GloVe (Stanford, 2014) FastText (Facebook, 2016) ELMo (Allen AI, 2018) ULMFiT (fast.ai, 2018) The Transformer Era Begins (2018–2020) GPT-1 (OpenAI, 2018) BERT (Google, 2018) Transformer-XL (Google/CMU, 2019) XLNet (Google/CMU, 2019) GPT-2 (OpenAI, 2019) — small, medium, large, XL RoBERTa (Facebook, 2019) ALBERT (Google, 2019) DistilBERT (Hugging Face, 2019) T5 (Google, 2019) — small through 11B CTRL (Salesforce, 2019) Megatron-LM (NVIDIA, 2019) ERNIE (Baidu, 2019) — ERNIE 1.0, 2.0, 3.0 Grover (Allen AI, 2019) XLM / XLM-R (Facebook, 2019) ELECTRA (Google, 2020) GPT-3 (OpenAI, 2020) — Ada, Babbage, Curie, Davinci GShard (Google, 2020) mBART (Facebook, 2020) DeBERTa (Microsoft, 2020) Reformer (Google, 2020) Longformer (Allen AI, 2020) BigBird (Google, 2020) Turing-NLG (Microsoft, 2020) BlenderBot 1.0 (Facebook, 2020) Scaling Up (2021) Switch Transformer (Google, 2021) — 1.6T parameters (sparse) GPT-Neo (EleutherAI, 2021) — 1.3B, 2.7B GPT-J-6B (EleutherAI, 2021) Codex (OpenAI, 2021) Jurassic-1 (AI21 Labs, 2021) DALL·E (OpenAI, 2021) — multimodal but LM-based Wu Dao 2.0 (BAAI, 2021) — 1.75T parameters PanGu-α (Huawei, 2021) HyperCLOVA (Naver, 2021) LaMDA (Google, 2021) Gopher (DeepMind, 2021) FLAN (Google, 2021) GLaM (Google, 2021) Megatron-Turing NLG (Microsoft/NVIDIA, 2021) — 530B WebGPT (OpenAI, 2021) Yuan 1.0 (Inspur, 2021) BlenderBot 2.0 (Meta, 2021) Cohere models (Cohere, 2021) The Explosion (2022) GPT-NeoX-20B (EleutherAI, 2022) InstructGPT (OpenAI, 2022) Chinchilla (DeepMind, 2022) PaLM (Google, 2022) — 540B OPT (Meta, 2022) — 125M to 175B BLOOM (BigScience/HuggingFace, 2022) — 176B Alexa TM (Amazon, 2022) — 20B UL2 (Google, 2022) Galactica (Meta, 2022) ChatGPT / GPT-3.5 (OpenAI, Nov 2022) Flan-T5 / Flan-PaLM (Google, 2022) Sparrow (DeepMind, 2022) Minerva (Google, 2022) AlphaCode (DeepMind, 2022) Jurassic-2 (AI21, 2022) Luminous (Aleph Alpha, 2022) text-davinci-003 (OpenAI, 2022) GLM-130B (Tsinghua, 2022) Tk-Instruct (Allen AI, 2022) Whisper (OpenAI, 2022) — speech, but transformer-based NLLB (Meta, 2022) — translation mT0 / BLOOMZ (BigScience, 2022) Stable Diffusion (Stability AI, 2022) — diffusion, but includes text encoder LMs YaLM (Yandex, 2022) — 100B 2023 — The ChatGPT Aftermath LLaMA 1 (Meta, Feb 2023) — 7B, 13B, 33B, 65B GPT-4 (OpenAI, Mar 2023) Alpaca (Stanford, 2023) — fine-tuned LLaMA Vicuna (LMSYS, 2023) Claude 1 (Anthropic, Mar 2023) Claude 2 (Anthropic, Jul 2023) Claude Instant (Anthropic, 2023) PaLM 2 (Google, May 2023) Bard (Google, 2023) — powered by PaLM 2 Gemini 1.0 (Google DeepMind, Dec 2023) — Ultra, Pro, Nano LLaMA 2 (Meta, Jul 2023) — 7B, 13B, 70B Chat variants Code Llama (Meta, 2023) Mistral 7B (Mistral AI, Sep 2023) Mixtral 8x7B (Mistral AI, Dec 2023) — MoE Falcon (TII, 2023) — 7B, 40B, 180B MPT (MosaicML, 2023) — 7B, 30B StableLM (Stability AI, 2023) Dolly (Databricks, 2023) RedPajama (Together AI, 2023) Phi-1 / Phi-1.5 (Microsoft, 2023) Orca / Orca 2 (Microsoft, 2023) Qwen 1 / Qwen 1.5 (Alibaba, 2023) Yi (01.AI, 2023) — 6B, 34B Baichuan (Baichuan Inc, 2023) — 7B, 13B ChatGLM / ChatGLM2 / ChatGLM3 (Zhipu AI, 2023) InternLM (Shanghai AI Lab, 2023) Xwin-LM (Xwin, 2023) DeepSeek (DeepSeek, 2023) — Coder, LLM, Math Zephyr (HuggingFace, 2023) OpenHermes (Teknium, 2023) Neural Chat (Intel, 2023) Starling (Berkeley, 2023) Puma / Platypus (various, 2023) WizardLM / WizardCoder / WizardMath (Microsoft, 2023) Gorilla (Berkeley, 2023) SOLAR (Upstage, 2023) Persimmon-8B (Adept, 2023) Inflection-1 / Pi (Inflection AI, 2023) Command / Command R (Cohere, 2023) Jamba (AI21 Labs, 2023) Nous Hermes (NousResearch, 2023) Fuyu-8B (Adept, 2023) BlenderBot 3 (Meta, 2023) SeamlessM4T (Meta, 2023) Segment Anything (Meta, 2023) — vision, but transformer-based Med-PaLM 2 (Google, 2023) Amazon Titan (Amazon, 2023) Jurassic-2 Ultra (AI21, 2023) Nemotron (NVIDIA, 2023) 2024 — Open Weight & Frontier Race Claude 3 (Anthropic, Mar 2024) — Haiku, Sonnet, Opus Claude 3.5 Sonnet (Anthropic, Jun 2024) Claude 3.5 Haiku (Anthropic, Oct 2024) GPT-4 Turbo (OpenAI, 2024) GPT-4o (OpenAI, May 2024) GPT-4o mini (OpenAI, Jul 2024) o1 / o1-preview / o1-mini (OpenAI, Sep 2024) Gemini 1.5 (Google, Feb 2024) — Pro, Flash Gemini 2.0 Flash (Google, Dec 2024) LLaMA 3 (Meta, Apr 2024) — 8B, 70B LLaMA 3.1 (Meta, Jul 2024) — 8B, 70B, 405B LLaMA 3.2 (Meta, Sep 2024) — 1B, 3B, 11B, 90B (incl. vision) LLaMA 3.3 (Meta, Dec 2024) — 70B Mistral Large (Mistral, Feb 2024) Mistral Small / Mistral NeMo (Mistral, 2024) Mixtral 8x22B (Mistral, 2024) Mistral Medium (Mistral, 2024) Codestral (Mistral, 2024) Pixtral (Mistral, 2024) Phi-2 (Microsoft, early 2024) Phi-3 / Phi-3.5 (Microsoft, 2024) — Mini, Small, Medium Qwen 2 / Qwen 2.5 (Alibaba, 2024) — various sizes Coder, Math, VL QwQ (Alibaba, 2024) — reasoning model DeepSeek-V2 (DeepSeek, 2024) DeepSeek-Coder-V2 (DeepSeek, 2024) Yi-1.5 (01.AI, 2024) InternLM 2 / 2.5 (Shanghai AI Lab, 2024) GLM-4 (Zhipu AI, 2024) Command R (Cohere, 2024) Aya 23 (Cohere for AI, 2024) DBRX (Databricks, 2024) Snowflake Arctic (Snowflake, 2024) Grok-1 (xAI, Mar 2024) — open weights, 314B MoE Grok-2 (xAI, Aug 2024) Gemma 1 / Gemma 2 (Google, 2024) — 2B, 7B, 9B, 27B RecurrentGemma (Google, 2024) PaliGemma (Google, 2024) Reka Core / Flash / Edge (Reka AI, 2024) Inflection-2.5 (Inflection AI, 2024) Jamba 1.5 (AI21, 2024) Nemotron-4 (NVIDIA, 2024) NVLM (NVIDIA, 2024) OLMo (Allen AI, 2024) — 1B, 7B StarCoder2 (BigCode, 2024) Stable LM 2 (Stability AI, 2024) Stable Code (Stability AI, 2024) Fugaku-LLM (RIKEN/Fujitsu, 2024) Sarvam-1 (Sarvam AI, 2024) — Indian languages SEA-LION (AI Singapore, 2024) EXAONE (LG AI Research, 2024) Solar Pro (Upstage, 2024) Granite (IBM, 2024) — various sizes MiniCPM (Tsinghua/OpenBMB, 2024) MAP-Neo (M-A-P, 2024) Sailor (Sea AI Lab, 2024) Tele-FLM (FlagOpen, 2024) Marco-o1 (Alibaba, 2024) 2025 (through May) Claude 3.5 Opus (Anthropic, 2025) Claude 4 Sonnet (Anthropic, 2025) Claude 4 Opus (Anthropic, 2025) Claude Sonnet 4.5 (Anthropic, 2025) Claude Opus 4.5 / 4.6 (Anthropic, 2025) Claude Haiku 4.5 (Anthropic, 2025) DeepSeek-V3 (DeepSeek, Jan 2025) DeepSeek-R1 (DeepSeek, Jan 2025) — reasoning model GPT-4.5 (OpenAI, 2025) GPT-o3 / o3-mini / o4-mini (OpenAI, 2025) Gemini 2.0 Pro / Flash (Google, 2025) Gemini 2.5 Pro / Flash (Google, 2025) LLaMA 4 (Meta, 2025) — Scout, Maverick Mistral Large 2 (Mistral, 2025) Grok-3 (xAI, 2025) Qwen 3 (Alibaba, 2025) Phi-4 (Microsoft, 2025) Gemma 3 (Google, 2025) OLMo 2 (Allen AI, 2025) Command A (Cohere, 2025)

> Do LLMs really understand…? > cutting edge models, eg StarCoder2 the crazy thing is that this is > Submitted on 18 Apr 2025 ie full 3 months after R1, not to mention o1-preview. Reasoners have changed everything. > University of Luxembourg ah. figures. too slow for sure

“Our testing approach includes 13 cutting-edge models, covering both code-specialized (e.g., StarCoder2) and general-purpose (e.g., GPT-4o) architectures” So obsolete models somehow show AI isn’t improving

We also rebuilt our chunker to be more concept-aware, deduplicated, and size-balanced. Research-backed evals show strong gains across multiple benchmarks: 4.3 points Recall@5 on RepoEval 4.3 points on CrossCodeEval for cross-language generation 2.67 points Pass@1 on SWE-bench 5.5 points average improvement with StarCoder2-7B on RepoEval

ok, so here's the tea. I tested: qwen3:14b, gpt-oss:20b, deepcoder:14b, deepseek-r1:14b, magistral:24b, starcoder2:15b, rnj-1 50% of the local models failed on test1. all but one got to test3. rnj-1 showed a lot of promise though. surprising. the one that got all the way was magistral:24b. unfortunately this doesn't fit in 16GB VRAM so it was extremely slow on the small (1-2%) that was CPU bound. Nevertheless it was OK. hosted models (i used ollama's free tier for this) did better with kimi-k2.5 looking the most promising. my primary beef now is that i still need to leverage lct-coder (my coding agent) for the best results. both opencode and codex had a very relaxed attitude toward the instructions file which isn't acceptable for lode coding where the lode coding prompt needs to be "font of mind" from the outset. so i now need to double-down on improving lct-coder in combination with kimi-k2.5 in order to liberate myself from claude code... i'm also going to see if i can squeeze rnj-1 for some smaller work on local, including some of the LCT AI features where larger models simply aren't necessary.

Efkrin's garage is starting to look like an AI version of Jay Leno's. Excellent. CodeLlama-70b-Instruct DeepSeek-R1-Distill-Llama-70B Dolphin-2.9.1-Llama-3-70B Llama-3.1-70B-Instruct Mixtral-8x22B-Instruct-v0.1 Qwen2.5-72B-Instruct Qwen2.5-VL-72B-Instruct Command-R-Plus Command-R-Plus-08-2024 DeepSeek-Coder-V2 Gemma-2-27B Hermes-3-70B Llama-3.1-70B Mistral-Large Phi-3-medium-128k-instruct Pixtral-12B-2409 SOLAR-10.7B-Instruct-v1.0 StarCoder2-15B Whisper-large-v3 WizardLM-2-8x22B

Tesla P40をOllama API経由で叩いたときの温度上昇確認。Grafanaで軽く可視化してみた。左から小さいモデル順。32bは長時間まわすとキツそう。 - llama3.1:8b - deepseek-coder-v2:16b-lite-instruct-q4_0 - starcoder2:15b - gpt-oss:20b - qwen2.5-coder:32b-instruct-q4_K_M

Tweet en forme d'aveu pour un dev old school réfractaire : je me suis mis au vibe-coding. 100% local. Backend Ollama avec Qwen3-coder:30b, starcoder2:3b, nomic-embed-text MCP server avec ma clé Perplexity Pro pour déléguer les recherches en ligne. Productivité x10 ¯\_(ツ)_/¯

맥미니M4에 사용할 java코딩용 오픈소스 LLM 추천 chatgpt : 아래는 Mac mini M4 기준 코드 품질 기준 Java 성능 기준으로 정리한 추천 순위입니다. 🥇 1순위: DeepSeek-Coder-V2 (16B 또는 33B) ➡ Java, Python, C , Go 등 전반적으로 현존 최고 오픈소스 코드 LLM 장점 Java 성능이 오픈소스 모델 중 최상위 코드 생성, 리팩토링, 테스트 코드 생성 모두 강함 오류 수정 능력(“debug reasoning”)이 뛰어남 압도적 코드 구조화 능력 영어 설명 → Java 변환 강함 메모리 기준 16B 모델 → M4 24GB~32GB에서 매우 안정 33B(Q4 quant) → M4 32GB에서만 가능 추천 실행 방식 ollama pull deepseek-coder-v2:16b 🥈 2순위: Qwen2.5-Coder 7B / 14B ➡ 한국어 Java 모두 잘함 / 다국어 코드 질문에 유리 장점 한국어로 Java 질문해도 오차 적음 코드 품질 안정적 디버깅 대화 능력 좋음 “왜 이렇게 되는지” 설명 능력이 강함 메모리 기준 7B → 16GB에서도 잘 돌아감 14B → 24GB 이상 추천 Ollama 명령 ollama pull qwen2.5-coder:14b 🥉 3순위: Code Llama 7B / 13B ➡ LLAMA 계열에서 코드에 강한 버전 장점 Java 코드 생성 정확도 준수 문서화·주석 생성 능력 좋음 익스텐션 개발 등에도 쓸만함 메모리 기준 7B → 16GB OK 13B → 24GB 추천 명령 ollama pull codellama:13b ⚡ 추가 추천 (특수 목적) • StarCoder2 (3B/15B) IDE 보조 목적, 완성(auto-complete) 목적이면 가볍고 빠름 GitHub-style Java 프로젝트에 최적화된 토큰화 • Codestral (Mistral AI) 코드 스타일 깔끔함 Mistral 계열 특유의 속도 강점 🏆 최종 결론 (Java 기준) 사용 목적추천 모델 최고 품질DeepSeek-Coder-V2-16B 한국어 JavaQwen2.5-Coder-14B 가볍고 빠름StarCoder2-3B or CodeLlama-7B 균형 좋은 범용 코드 모델Qwen2.5-Coder-7B ✨ “M4 16GB/24GB/32GB 기준” 정리 🇳🇱 16GB StarCoder2-3B Qwen2.5-Coder-7B CodeLlama-7B 🇳🇱 24GB Qwen2.5-Coder-14B DeepSeek-Coder-V2-16B 🇳🇱 32GB DeepSeek-Coder-V2-16B (가장 추천) DeepSeek-Coder-V2-33B(Q4 양자화)

if you have heard of GPT-4o, o1-preview, o1-mini, GPT-4 Turbo, GPT-3.5, Claude 3.5 Sonnet, Claude 3 Opus, Claude 3 Haiku, Claude 2.1, Gemini 1.5 Pro, Gemini 1.5 Flash, Gemini Ultra, Gemma 2, Gemma, Llama 3.2, Llama 3.1, Llama 3, Llama 2, Llama 1, CodeLlama, Mistral 7B, Mistral Large, Mixtral 8x7B, Mixtral 8x22B, Codestral, Ministral, DeepSeek-V3, DeepSeek-R1, DeepSeek Coder, Qwen 2.5, Qwen 2, Qwen-VL, Command R, Command R , OLMo, OLMoE, Phi-3.5, Phi-3, Phi-2, Orca, Grok-2, Grok-1, DBRX, Jamba, Yi-Large, Yi-1.5, Falcon 180B, Falcon 40B, Solar 10.7B, Nemotron-4, StableLM, Stable Beluga, WizardLM, Vicuna, Alpaca, Guanaco, Koala, Zephyr, StarCoder2, Pythia, Nous Hermes, Ziphu, Kimi, MPT-7B, MPT-30B, Bloom, OPT-175B, PaLM 2, PaLM, Gopher, Chinchilla, Switch Transformer, WuDao 2.0, XLNet, BERT, RoBERTa, T5, Ernie Bot, Titan, PLATO-XL, Doubao-Pro, Doubao, Kimi k1.5, Kimi, Moonshot, Hunyuan-Large, Hunyuan, HunyuanVideo, Baichuan 4, Baichuan 3, Baichuan 2, Baichuan, MiniMax abab6.5, MiniMax, InternLM 2.5, InternLM2, InternLM, SenseNova 5.0, SenseNova, SenseChat, Spark Desk 4.0, Spark Desk, Pangu 5.0, Pangu-Sigma, Pangu-Alpha, Wu Dao 3.0, Wu Dao 2.0, Skywork-MoE, Skywork, Aquila2, Aquila, CPM-Bee, CPM, MOSS, XuanYuan, BlueLM, Taiyi, Yuan 2.0, Linly You are too invested in LLMs.