Bielik-One explained: four open Polish decision models
Bielik-One is SpeakLeash's first decision-model family: four Apache-2.0 checkpoints, 4B to 11B, returning probabilities. Specs, access and what is missing.
Releases
Every new model we spot gets a post: a tested review when we can run it, a clearly labelled first look when we can't yet. LLMs, decision models, voice, image and video.
Bielik-One is SpeakLeash's first decision-model family: four Apache-2.0 checkpoints, 4B to 11B, returning probabilities. Specs, access and what is missing.
Perplexity's Decider v1.3 is a 26B Apache-2.0 decision model on Qwen3.8-27B that adds Locate span extraction to choice, yes/no and score. Specs and claims.
Protegrity-PII is an open-weights XLM-RoBERTa model that tags 69 PII types in 7 languages. Specs, the vendor-reported scores, and where the claims stop.
SOV-Transcribe-1 is an Apache 2.0 speech-to-text model with 4.6B active parameters and 7 languages. Specs, vendor-reported scores and the limits to know.
Baidu's ERNIE Wizshot Micro-Expression turns text, images and speech into expressive video, with 4,000-character inputs on Qianfan. What is known, what is not.
Translated's Lara Mini is a 7.2B mixture-of-experts base model with 0.6B active parameters, trained on 45 languages. Specs, claims, license and who should care.
We tested inclusionAI's free Ling 3.2 Flash against Ling 3.1 Flash and DeepSeek V4.1 Flash on 8 dev tasks. Quality is close; latency is 10x worse.
LiquidCodec is Liquid AI's 74M-parameter streaming speech codec behind LFM3-Audio-2B. Specs, speed claims versus Mimi and NanoCodec, access, and who it suits.
Retell says its own Tell-1 ranks first on its vCX voice benchmark, leading on colloquial speech and latency. What is claimed, what is missing, who should care.
Tether's BotPsy-0.8B scores 44.0 on BFCL v4 versus 23.4 for its Qwen3.5-0.8B base, beating a model 2.5x larger on tool use. Claims, limits, who should care.
Chariot-1-TTS-Multilingual is a text-to-speech model for 11 Indian languages plus English, with 50+ voices and code-mixing. Specs, pricing and claims.
We ran Gemini 4 Hydrogen against Gemini 4 Carbon and GPT-6.1 Terra on 10 coding-agent prompts. Mostly ties, but it is 2.5x faster than Carbon at $1.50/$7.50.
ByteDance's Seedance 2.5 Flash makes video up to 30 seconds at about 0.99 yuan per second in 720p, 34% below Seedance 2.5. Specs, price per clip, API date.
TIER IV open-sourced five end-to-end driving models for Autoware on 2026-11-11, from 16M-parameter METEOR to a 2B NVIDIA model. What each does, who it fits.
DeepL Aure is a voice AI model that adapts translated speech to a speaker's tone, emotion and pace. Languages, access, what DeepL claims and what is untested.
Google's Gemini Decision 1 scores up to 1,000 questions on one input for $0.10 per million input tokens and no output charge. Specs, claims, who should care.
Gemini Robotics 2.5 is DeepMind's newest robot-control model. What it reports against version 2, who can use it today, and why the numbers are vendor-only.
PAN 2 is MBZUAI's open-weight world model that turns a scene and an action into video. Specs, license, hardware, vendor benchmarks and what is still unverified.
inclusionAI's SingVL_S1_4B is an open-weights model on Qwen3.5 that returns probabilities for choices, yes/no and scores over text and images. What it is.
Zyphra's ZONOS3-Mini is an Apache 2.0 text-to-speech model with 800M total and 100M active parameters, voice cloning and a browser build. Specs and claims.
We ran Baidu's ERNIE 5.5 Flash Preview ($0.28/$1.13 per 1M tokens) against ERNIE-Light-DT, DeepSeek V4.1 Flash and Qwen3.7 Plus on 8 real tasks.
We ran Gemini 4 Carbon against Argon and Claude Fable 5.2 on 8 hard engineering prompts. It won 5-0 and 2-2 at $2/$10 per million tokens, doubling after Nov 30.
We ran Meta's Muse Ember against Claude Haiku 5.5 and GPT-6 Luna on 8 agent tasks. It beat Haiku 4-0 but cost 6.6x more per answer than Luna.
We tested Nex-N2.6 ($0.12/$0.40) against N2.5-Pro, GLM-5.3-Flash and DeepSeek V4.1 Flash on 8 agent tasks. It beat its GLM base 3-0 and lost to DeepSeek 0-3.
boost.ai's Presis is a fine-tuned, self-hosted model that makes routing, scope and safety calls in AI agent chats. What it does, the claims, who should care.
Thalamus Decision is Ingenix's open-source 9B model for aging biology. What the company reports on LongevityBench, what is missing, and who should look first.
We ran Claude Fable 5.2 against Fable 5.1 and GPT-6 Astra on 8 real dev and support prompts. It won 6 and lost none, at $8/$40 per million tokens.
GR00T 2 is Nvidia's open world action model for humanoid robots. What Nvidia reports, what is still missing, and who should look at it first.
K-EXAONE 2.5 is LG's 750B model with a free OpenAI-compatible API. What LG reports on agentic benchmarks, what is missing, and who should try it.
Empero Manta claims a 100M-token prompt. Early access only, self-reported tests, no price or API. What is claimed, what is missing and who should care.
Gemini Omni 1.2 Flash makes video with audio up to 15 seconds at about $0.07 per second in 720p, 30% below Omni 1.1. Specs, price per clip, who should switch.
Ginza 1.0 is De¹'s waitlist-only financial decision model that forecasts how an on-chain trade will execute before you sign. What is known and what is not.
GPT-6 Cyber is OpenAI's gated security model at $20 per million input tokens. What OpenAI reports, how the price compares, and who can get access.
Salesforce's Koa Decisions is a Gemma-based decision model that picks an answer from text, images or audio in one step. Specs, vendor claims, access and gaps.
S5-nano is Maritaca AI's 4B Apache-2.0 vision-language model tuned for Portuguese. Specs, vendor-reported scores vs Qwen3.5-4B, caveats and who should care.
Arete-27B is Nous Research's Apache-2.0 fine-tune of Qwen3.8-27B, built to cut state-propaganda answers. Specs, vendor-reported evals, caveats, who should care.
Video Rebirth's HyperFlow H3 Flash shrinks MiniMax-H3 from about 31B to 13.7B parameters, makes video and audio in 4 to 8 steps, and runs on one RTX 4090.
Inkling 2 Small is Thinking Machines' 594B MoE with 17B active parameters, live on Tinker with Apache 2.0 weights promised. Specs, vendor evals and caveats.
Mirage is a 51.3M-parameter open vision model that describes, detects, reads text and segments in one pass, in 22 MB. Specs, vendor numbers and caveats.
Underdog Woof is a 38M-parameter English text-to-speech model that runs on CPU in a 31.7 MB download. Specs, vendor-reported WER, license limits, who it suits.
Wizstar's Wizshot 1.0 is a reference-to-video model: up to nine input images, 720p clips of 5 to 15 seconds, 5 credits per second. Specs, claims and caveats.
169Pi's Alpie Seed is a 9B Apache 2.0 model built on Qwen3.5-9B that runs offline. Specs, vendor benchmark claims, hardware needs and who should care.
Kakao's Kanana-2-v-decide is a 3B open-weight decision model that returns probabilities instead of text. What it does, its claims, and who should care.
Why LLMs hallucinate, the three error types, prompting patterns that reduce them, and the one production eval that catches failures before your users see them.
JSON mode vs structured output: how constrained decoding works, when 3% failure rates become a real cost, and which enforcement approach your pipeline needs.
A system prompt is the hidden instruction block that shapes every LLM response. Here's how it works, what it costs, and the production pitfalls to avoid.
Fine-tuning trains a model on your data to shape behavior and output format. Here's when it beats prompting, what it costs in 2026, and when to skip it.
Function calling lets LLMs trigger your code via JSON schemas. How the loop works, when to use it over plain prompting, and the pitfalls most builders hit.
Embeddings convert text into vectors that capture meaning. Here's how cosine similarity works and how to pick the right model for your RAG pipeline.
Prompt caching cuts LLM API costs up to 90%, but Anthropic, OpenAI, and Gemini implement it differently. Here's how each vendor's billing actually works.
Claude Fable 5 review with real benchmark data: 5-3 over Opus 4.8, 3-0 vs GPT-5.5 on 12 coding and reasoning prompts. Includes subscription break-even math.
We ran 12 coding, math, and data tasks through Opus 4.8, Opus 4.7, and GPT-5.5. Opus 4.8 swept GPT-5.5 but split with its predecessor.
We ran 5 developer tasks through DeepSeek V4 Pro, GPT-5.5, Opus 4.7, and Llama 4. V4 Pro beats GPT-5.5 while costing 4.5x less, but latency averages 28 seconds.
OpenAI's GPT-5.5 brings a 1M-token context and native computer use to the frontier, at double GPT-5.4's price. Here's what actually changed.
The context window is your LLM's working memory per call. What 128k tokens actually fits, why usable size is smaller than advertised, and how to check yours.