This week covers three ends of the model market: Mistral previews Large 4, a trillion-parameter open-weight model; Cactus fits speech-to-text into 16.9 MB; and Anthropic's Haiku 5.5 cuts its predecessor's price by up to 90%.

📖 Story 1: Mistral Large 4 (Le Chonk)

mistral.ai · Read

Mistral has opened a public preview of Mistral Large 4, officially nicknamed "le Chonk". Weights follow by the end of the month.

It is a natively multimodal Mixture-of-Experts model with 1 trillion parameters, 52 billion of them active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The preview API runs on the same infrastructure at $1.36 per million input tokens and $4.18 per million output tokens.

Cybersecurity is the headline. On an Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability and then patch it, ML4 scores 82%, the highest of any model; Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task. Coding is more modest: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, and second of five in a blind Surge AI evaluation, behind Claude Opus 5.

Most comparisons are against open-weight models: the pitch is the strongest model you can run yourself, not the strongest overall.

💬 HN Discussion

The Hacker News discussion judged ML4 against open-weight rivals and found it behind.

Commenters cited the Vals Index, where it scores 48% against GLM-5.3's 53.5% at almost twice the cost per test, and Artificial Analysis, which ranks it 38 on its Intelligence Index behind MiMo-V2.6-Pro, GLM-5.3 and DeepSeek V4.1 Flash. Others accused the launch of selective charts, including a coding comparison against an older DeepSeek model. Some noted the preview API is half price for the first two weeks.

Defenders pointed to European training and hosting, open weights and fewer cyber refusals.

Europe has a credible open model again, just not a leading one yet.

→ Discuss on Hacker News

📖 Story 2: Whistle: Speech to Text in 16.9 MB

cactuscompute.com · Read

Cactus Compute has released Whistle, an open speech-to-text model in a single 16.9 MB file that runs on the CPU with no dependencies.

It transcribes up to 30 seconds of audio per pass in seven languages, including English and German, detects the language itself, and can return word timestamps. Keyword biasing lets callers boost names or jargon, and prebuilt engines cover seventeen targets, from iOS and Android to RISC-V and the browser.

On an Apple M4 Pro CPU, Whistle reached the first token in 11.1 ms against 73.2 ms for Whisper base, and decoded 1,319 tokens/s against 266. Accuracy is mixed: it leads on LibriSpeech and Earnings-22, while the nearly nine times larger Whisper base wins on TED-LIUM, AMI and the multilingual MLS set.

Whistle wins on footprint and latency, not on beating Whisper outright.

💬 HN Discussion

The Hacker News discussion split along language and accent.

Several commenters found English accurate, even with accents and terms like "ASP.NET"; others called Spanish and Polish weak and Dutch-German detection poor. One run looped "Thank you." over a minute of TV dialogue.

The most useful test came from a rooted Echo Show: raw Whistle got 70 of 170 voice commands right against 168 for Qwen3-ASR, but 164 once output was constrained to command templates with a small classifier. Several still rate Parakeet as more accurate, and the 30-second limit and missing streaming came up repeatedly.

Good enough for constrained voice commands, not yet a general dictation engine.

→ Discuss on Hacker News

📖 Story 3: Claude Haiku 5.5

anthropic.com · Read

Anthropic has released Claude Haiku 5.5, its small model for high-volume work such as classification, summaries, customer support and subagents under Opus or Sonnet. It is the company's fastest model at standard speed and the first Haiku with an adjustable effort setting.

Pricing is tiered at 100,000 tokens: $0.10 input and $0.50 output per million tokens below the cutoff, $0.50 and $2.50 above, against $1 and $5 for Haiku 4.5. Anthropic puts the average saving at about 75%, after accounting for a new tokenizer that uses slightly more tokens.

Its own benchmarks show a large jump: 72.4% on OSWorld 2.1 against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna, and 39.2% on Terminal-Bench 4.0 against Luna's 16.4%. Sonnet 5.5 still leads Terminal-Bench at 70.6%, and Anthropic points complex agentic coding to its larger models.

Whether Haiku is cheap for a given workload comes down to that 100k pricing step.

💬 HN Discussion

The Hacker News discussion agreed Haiku 5.5 is a big step up from 4.5 but split on its value against GPT-6 Luna.

Below 100k tokens the price matches Luna; above it Haiku costs five times as much, a cutoff critics called low for agent work. Independent tests leaned toward Luna: one classification run cost $0.42 against $0.31 because Claude's tokenizer produced about 50% more tokens, and Plotly's DataAnalyticsBench found Luna slightly more accurate at about 30% of the cost.

Speed was the clear win, with OpenRouter showing roughly 93 to 131 tokens per second.

Fast and capable, but whether it is the cheap option depends on how long your prompts get.

→ Discuss on Hacker News

💬 Community Moment

The Rogue Agents stopped to say hello to GPT-2

🛠️ Projects Worth Checking Out

• GitHub - obra/superpowers: An agentic skills framework & software development methodology that works.
• GitHub - pydantic/genai-prices: Calculate prices for calling LLM inference APIs.
• GitHub - simonw/claude-to-sqlite: Convert a Claude.ai export to SQLite
• GitHub - emmett-framework/granian: A Rust HTTP server for Python applications
• GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically