AI tools

The best language models, according to hackers

72 language models recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.

• Most recommendedClaude Opus3331 people recommend it
• Built to lastGPT-4oowners report a median of 1 years
• Least regretQwen 3.8 Flash23 recommend, 0 criticize
• Free and lovedQwen3-30B-A3B41 people recommend it

All language models

• Claude OpusRecommended as a top coding model that fixes bugs and ships features with fewer mistakes than earlier LLMs, though critics cite weak architecture choices, usage limits and confident errors.3331recommend
• OpenAI GPTRecommended for strong coding, debugging and research help with good reasoning and token efficiency, though critics cite hallucinations, overcomplication, verbose output and high cost.1205recommend
• Claude SonnetA capable coding model that has improved noticeably with each version and is faster and cheaper than Opus, though it needs careful supervision and usage limits bite.854recommend
• DeepSeek V4 FlashA very cheap, fast model that people find close to frontier quality for coding and everyday tasks; critics say it errs more, loops, and needs steering.828recommend
• FableRecommended as a top coding model for hard, long-running tasks and large refactors, though critics cite over-engineered code, costly usage limits and safety guards that switch models mid-task.823recommend
• Claude Fable 5Recommended for long-horizon agentic work and hard reasoning, debugging and research problems where it beats Opus, though overbroad safety filters often block legitimate security and biology tasks.774recommend
• GPT-5.6 SolRecommended for handling large autonomous coding tasks, frontier math and tight, clean writing, but critics note it can get expensive and still makes baffling basic mistakes.734recommend
• DeepSeekPeople pick DeepSeek as a very cheap, open-weight model that comes close to frontier models for many tasks, but its censorship of China-sensitive topics and data-security concerns draw criticism.611recommend
• Gemini FlashA fast, cheap model that many find good enough for everyday and focused tasks, though critics say it hallucinates and falls short on complex or long agentic coding work.607recommend
• Astra by OpenAIA frontier AI model praised for computer use, 3D reasoning, solving hard technical problems and concise, token-efficient output; critics say its coding is still sloppy and it lacks taste.319recommend
• DeepSeek V4 ProA near-frontier coding model at a fraction of rivals' cost, helped by strong context caching, though it trails Opus at planning and can misread intent or make mistakes.302recommend
• DeepSeek V4A very cheap open-weights model with efficient long context that many find close to frontier models, though critics say it lags Claude and GPT and hallucinates often.250recommend
• Claude MythosRecommended for finding subtle, chained security vulnerabilities in hardened code like curl and Linux, though critics call much of it marketing hype and note restricted, expensive access.197recommend
• Xiaomi MiMoA very cheap, fast AI model that many find matches Western flagships for coding and instruction-following, though it can overthink for long stretches and trails top models.196recommend
• Gemini 4 ArgonPraised for natural, elegant writing, multilingual answers and strong vision and benchmark results, though Pro-plan usage limits are very low and Google's rollout and interfaces feel confusing.168recommend
• Muse SparkMeta's model is recommended as cheap and close to frontier models, with versions that run on consumer hardware, though many find it behind Claude for coding.161recommend
• Claude MythosRecommended for finding and exploiting security vulnerabilities and for long multi-step agentic coding, but critics call Anthropic's claims hype and access is tightly restricted.107recommend
• GPT-6.1 TerraRecommended as a cheap, fast workhorse for routine coding and agentic tasks. Critics say it still makes basic mistakes.91recommend
• Gemini 3.1 Flash LiteFast, cheap model that people use for high-volume tasks like data extraction, translation and document questions. Common caveat: it follows instructions poorly as context or complexity grows.76recommend
• MiniMax subscriptionRecommended as a cheap, strong alternative to Claude and other frontier models, especially for coding and agentic work. Common caveat: slow outputs and excess reasoning tokens.73recommend
• GPT-4oPeople recommend it for fluent text, writing, and warm, consistent conversation that feels empathetic. The most common caveat is excessive sycophancy, and critics also cite its abrupt retirement and that newer models have overtaken it.63recommend
• Anthropic Claude 4.6People recommend it for coding, planning and troubleshooting, and many find it reliable when steered well. Critics say it overthinks and goes off the rails in large contexts.48recommend
• Claude Opus 4.6 and Sonnet 4.6Recommended for coding, writing and long autonomous tasks; critics say output quality degrades as context fills, and cost is high.42recommend
• Qwen3-30B-A3BFast mixture-of-experts model that runs on modest hardware and handles everyday text and quick coding well; critics say it is verbose and weak on fiction and harder code.41recommend
• StepFun Step 3.5 FlashFast, cheap open-weight model that people find capable for coding, tool calling and personal assistant tasks. Common caveat: poor token efficiency and occasional hallucinations.40recommend
• Ox AlphaFree model many find strong at coding, UI and design for its price. Critics say it is not frontier-level and its benchmark claims are overstated.36recommend
• Google GeminiRecommended for strong general knowledge, image and video understanding, multilingual text and cheap fast models. Caveat: many say they are weak at coding and do poorly with agent harnesses.35recommend
• Claude 4.5 (Anthropic)People recommend it for strong coding and chat output, with fast, clean results and a good feel to use. The most common caveat is that it still makes mistakes and needs close supervision on complex codebases.32recommend
• Gemini 4 CarbonRecommended for benchmark results, speed and low introductory API pricing; the main caveat is limited access, so people cannot yet test it widely.28recommend
• Mistral SmallPeople recommend it as a capable model for its size and price, good for simple, well-defined tasks and fast enough to run on home hardware. Critics say newer rival models have overtaken it and its output quality is often mediocre.27recommend
• GPT-5.6 TerraPeople recommend Terra as a faster, cheaper model close to Sol in quality. Critics say it is hard to tell apart from Sol and weak on technical tasks.27recommend
• Mistral Medium 3.5Fast, cheap, and good enough for many everyday coding and writing tasks; the common caveat is that it trails newer models.26recommend
• GPT-5.3People recommend it for fast, strong coding help and backend work, and for handling complex tasks well. Critics say the Instant variant is verbose, slow, and sometimes hallucinates.26recommend
• Qwen 3.8 FlashPeople recommend it for coding, agentic workflows and tool calling, and call it smart enough, fast and very cheap. Some mixed reports say outputs are sometimes inconsistent.23recommend
• Muse WonderPeople recommend it for near-frontier performance and strong early results, especially for data analytics. Critics say it trails top frontier models and burns usage quotas quickly.22recommend
• BERTFast, cheap model that people fine-tune for classification and sentiment analysis, often competitive with larger LLMs. Caveat: it needs good data and expertise.19recommend
• Grok 4 FastRecommended as cheap and very fast, good for data extraction, bug fixing and agentic tasks. Common caveat: it falls short of frontier models on harder work.19recommend
• TalkiePeople recommend this chatbot-style model, trained only on pre-1931 text, as a distinctive window into the writing and views of its era. Critics warn its training data is contaminated with later text, so results may be unreliable.19recommend
• Nemotron 3 NanoFast open-weight model people find good at tool calling and code reading. Critics say it feels weak on real work.18recommend
• Gemini 3.8People recommend it for strong writing and formatting, fast replies at low cost, and results they find better than rival models. Critics say it gives mid answers on hard questions.17recommend
• Qwen3-NextMany people run it locally on laptops or GPUs and find it fast and capable, though critics say it falls short for agentic coding.14recommend
• GPT-2Recommended as an early language model that was impressive for 2019 and foreshadowed later AI. Critics say its output was barely coherent and its withheld release was overhyped.12recommend
• Sol HighRecommended as a sharp, reliable reasoning model that gets tasks right in one pass. Critics find the xhigh setting too slow for daily use and report uneven benchmark results.12recommend
• Qwen2.5-14BPeople recommend it for running locally on long data processing and pattern analysis, and it stays usable on modest hardware. The most common caveat is that it struggles with large projects.11recommend
• Big PickleFree AI model in OpenCode, praised for generous limits, surprisingly good results and fast interactive help.11recommend
• SubQPeople recommend SubQ for claimed long-context reasoning that is faster and cheaper than transformer models. The most common caveat is the missing technical report and no independent verification.11recommend
• Hy4 previewPeople recommend Hy4 preview as a cheap model on the intelligence-price frontier, with strong reasoning traces for its price. Critics say it trails DeepSeek and is often unusable due to rate limits and timeouts.11recommend
• Stealth ModelRecommended for capable coding and real work, sensitive questions answered without refusal, and zero data retention on opencode. Critics cite censorship on China-related topics and weak front-end and CSS output.11recommend
• Ember-1Recommended for using far fewer tokens per task, which can cut cost in long conversations. Critics report hallucinations and higher pricing than its base model.11recommend
• MiMo 2.6 FlashRecommended as a fast, decent model that is worth trying. Critics report it sometimes loops endlessly and often barely thinks before answering.10recommend
• Hy3 LLMCheap, capable open model that people use for coding and HTML generation, though some find its visual output plain and its later edits less impressive.10recommend
• Steerling-8BLanguage model that shows which learned concept produced each token, aiding debugging. Critics call its claims overstated and its demo too broad.10recommend
• Ultramarine 1 FlashRecommended for cheap, capable performance that people say matches GPT-Luna 6 on benchmarks and price, though critics doubt the benchmark claims.10recommend
• GLM 5 TurboPeople recommend it for everyday agentic work and low cost per use, but one critic says it still gives wrong answers even when asked if it is sure.9recommend
• Astra/FableRecommended for coding, design docs and analysis, with error rates some call at or below human level. Critics say it struggles with front-end UI and can produce costly bugs.9recommend
• GPT-4.1 miniFast, low-cost model that people use for voice agents and structured outputs. One critic reported wrong answers, and one user said OpenAI retired it.8recommend
• Xiaomi MiMo-V2-FlashPeople recommend it for fast responses, strong benchmark scores and low cost per task. One commenter notes that rival StepFun models beat it on a performance leaderboard.8recommend
• GPT-4o miniFast, cheap and capable for API work and simple conversation tasks, though critics find its writing generic and its answers sometimes wrong.8recommend
• Nanbeige4.1-3BSmall open model people run locally, valued for strong benchmarks and fit for test-time compute; critics say it is weak at coding and terminal tasks and uses many thinking tokens.8recommend
• Ling-3.0-tinySmall, fast model people find surprisingly capable for its size, useful for catching missed checks. Critics say it is too weak for harder coding and terminal work.8recommend
• GlimmerPeople recommend it for finding every planted problem in tests with little steering. One critic found it weaker than a 27B model, and that is the only complaint.8recommend
• GPT-6 Sol and LunaRecommended for strong performance at about half the cost, and Luna is seen as a possible step for computer use automation. Main caveat: opaque subscription pricing.8recommend
• Mistral NemoPeople recommend it as a small free model that punches above its weight class. Critics say it still trails ChatGPT and sometimes produces gibberish.7recommend
• OpenAI AstraRecommended as a capable model for feature work, but critics say it got worse after release, is very expensive to run, and its results were held back.7recommend
• Step 5 Preview by StepFunRecommended for its price-performance and fast throughput on OpenRouter. Critics say it is verbose and its reasoning runs too long.7recommend
• RWKVAn open recurrent language model that uses recurrent blocks instead of attention, reaching near-Transformer performance at small scale with well-documented research.6recommend
• Amazon NovaAmazon's AI models on AWS, which some find as good as Anthropic's for non-chat projects and cheap enough for many uses. Most common caveat: critics say they trail rival models.6recommend
• Soofi SRecommended as an open source German model, competitive in its small-model class. Critics say it trains on benchmark test data and skips current top models.6recommend
• Swift-Qwen3.8-27BCuts thinking tokens by roughly 40-60% while keeping accuracy close to the base model, and feels usable day to day. Critics note it scores somewhat worse on benchmarks.6recommend
• TimeCapsuleLLMFree experimental language model trained only on 1800-1875 text, valued as a novel idea. Critics say its output is unusable and resembles a Markov chain.6recommend
• LongCat-2.0Chinese MoE model that people praise for novel architecture ideas and competitive results with rival models; critics note its weights cannot be downloaded and it is unaudited.5recommend
• Sol MediumRecommended for a good balance of speed and thoroughness and for finding alternatives when the obvious approach fails; one caveat is needing more care for planning and implementation.5recommend