Text-to-speech in 64.5 MB.
2 patches before the weekend.
An agent that writes its own tools.
▲ AI DEV SIGNALS
by Marktechpost · October 23, 2026
Hi there,
Your agent can read the web for $0. TinyFish Search and Fetch are free with no credit card, and you pay only when your agent acts. Top up before Oct 31 and get 30% extra. (Sponsored)
And now, here's what moved in AI dev this week. Gradium open-sourced a 31M-parameter speech model that starts talking in 7 ms on a laptop CPU. Cohere's new reranker cuts agent tokens by up to 41%. Liquid AI claims LLM answers in under 50 ms. Mistral shipped a full voice agent stack.
01 · Your Raspberry Pi just learned to talk
Gradium, a Paris voice AI company spun out of Kyutai, open-sourced Phonon, a family of on-device text-to-speech models. The English Phonon 31m ships under Apache 2.0. It needs no cloud call, no GPU, and no per-character bill.
How it works: A causal transformer reads raw text through a 4,000-token SentencePiece vocabulary, with no phoneme step. It generates 1 latent vector per 80 ms frame. A flow-matching head turns each vector into audio in a single step, and a small Mimi decoder outputs 24 kHz audio. Phonon 31m is 6 layers wide at 512, distilled from a larger teacher.
The numbers: 1.70% WER on 500 hard sentences full of phone numbers and emails, against 3.50% for Kokoro and 8.71% for NeuTTS Nano. First audio arrives in 6.9 ms on a MacBook Air M5 CPU, against 517 ms for Kokoro. On a Raspberry Pi 5, it starts in 38 ms and runs 3.7x faster than real time. In blind listening tests, it scored 2098 Elo to Kokoro's 2043.
The fine print: Only the English 31m model is Apache 2.0. The 40m and multilingual models are CC BY-NC 4.0, so they're non-commercial. On the easier Seed-TTS set, Kokoro edges ahead, 0.45% to 0.51%. All benchmarks are Gradium's own.
02 · 4 more drops worth stealing for your stack
It comes in Pro ($0.05 per 1M tokens) and Fast ($0.02) tiers, with 32K context and 100+ languages. A new Context Focus feature returns the relevant passages inside each document. In Cohere's search-agent test, that cut tokens 35% against BM25 alone and lifted accuracy from 46% to 58%. Scores are calibrated, so 0.7 means a 70% chance the document is useful. All numbers are self-reported, and Pro's median latency is 11.7 s per 96-document request.
Datalab released per-test verdicts for 27 document parsers, and every number on the leaderboard can be rebuilt from them. The scorer is open code that calls no model. Grade any parser on your own PDFs with 1 score call. Benchmark your parser. (Sponsored)
RIPL is a serving engine built for live request paths like search ranking and fraud checks. On 1 B200, LFM2.5-1.2B hit 34.6 ms p95 at 1,000 req/s, against 498 ms for Qwen3.5-0.8B on vLLM. Shopify already runs LFMs at sub-20 ms in production search. RIPL is enterprise-only and closed, and the 50 ms target slips on longer requests.
The release includes 3 transcription models, 1 TTS model, and a LiveKit-based SDK with sub-second turns. Small Transcribe 3 ranks #3 on Artificial Analysis at 2.1% WER and leads Mistral's diarization chart at 18.4% cpWER. Batch transcription costs $0.0035 to $0.004 per minute. The trade-off: the realtime model trails Gemini 3.5 Realtime on accuracy, and there are no open weights.
03 · Signals: 2 patches and an agent that builds its own tools
Simular's Agent S4 writes its own tools. A meta-agent reads task runs, with no success labels, and rewrites the Python interfaces the agent uses to drive a desktop. With Qwen3.8-27B, it hits 21.3% on OSWorld V2, 3.8x the same model with fixed interfaces. That beats Claude Opus 4.8's 20.6% at a fraction of the cost. Code is open:
pip install agent-s4.Patch Node.js today. The Oct 21 release fixes 10 flaws, 4 rated high. The worst is an out-of-bounds write in the bundled Brotli decoder that allows remote code execution. Upgrade to 22.23.4, 24.22.1, or 26.12.1; Node 20 and older get no fix.
State hackers ran an open-source AI agent framework. Microsoft says China-aligned Storm-0940 used ARTEX after the ConnectWise breach. It tried to sign in to more than 2,250 Microsoft 365 tenants, peaking at 2,900 attempts per minute.
8 days left on 30% extra. TinyFish Search and Fetch are free for every agent, with no card required. Top up your Wallet before Oct 31 and get 30% on top. (Sponsored)
PyTorch runs natively on AWS Trainium. TorchNeuron is an Apache 2.0 PyTorch backend in public beta. Swap
.to('cuda')for.to('neuron'), and Transformers, Diffusers, and TRL run with no Neuron-specific code. Deepgram reports about 60% more audio processed per dollar than its GPU baseline. The beta wants static shapes.OpenAI's Decisions API is now GA. It costs $0.10 per 1M input tokens with no output charge and takes up to 400 questions per request. OpenAI says it's up to 10x faster than GPT-6 Luna through the Responses API.
Can your model write COBOL? OpenAI open-sourced MultiLangBench, which has a model build a SQLite-compatible database engine in 20 languages. GPT-6 Astra averages 87.4%, and COBOL is the hardest at 72.1%. These are OpenAI's numbers on its own models, but the test suite is on GitHub to run yourself.
From our partners
OmniParseBench grades document parsers with 16,288 pass/fail checks across 2,937 pages, 92 languages, and 22 suites. Every verdict is public, so you can check the rankings yourself, then run the same tests on your own documents. (Sponsored)
1 question
If your text-to-speech fits in 64.5 MB and runs on a Raspberry Pi, which voice feature are you still sending to the cloud?
Have a great weekend.
until next week,
the marktechpost team
REACH 70K+ AI BUILDERS
Was this email forwarded to you? Sign up here