AI Daily Brief — 13 March 2025
A Thursday driven by enterprise-grade efficiency stories. Cohere shipped Command A with a deliberately small two-GPU footprint, Sesame open-sourced the speech model behind viral assistant Maya, and Insilico Medicine turned its AI drug-discovery pipeline into another unicorn raise.
Top stories
- Cohere releases Command A — 111B params, 256K context. Dense model with 23-language support and 150% higher throughput than Command R+ 08-2024, runnable on just two A100/H100 GPUs. Cohere claims it matches or beats GPT-4o and DeepSeek-V3 on agentic enterprise tasks at up to 156 tokens/sec. SDK identifier: command-a-03-2025. via Cohere
- Command A’s enterprise efficiency story. Launch coverage framed it as ‘max performance, minimal compute’ — two-GPU footprint vs larger rivals, 50% inference-cost reduction, with multilingual NTREX score of 68.8 vs GPT-4o 69.8 and DeepSeek-V3 71.0. Throughput claimed at 1.75x GPT-4o and 2.4x DeepSeek-V3. via AIwire
- Sesame open-sources CSM-1B under Apache 2.0. The startup behind viral assistant Maya released the base conversational speech model on Hugging Face and GitHub. 1B-parameter model with a Llama backbone plus a smaller audio decoder producing Mimi codes, trained on ~1M hours of English audio; can clone a voice from ~1 minute of source. via TechCrunch
- Insilico Medicine closes $110M Series E. AI-driven drug-discovery company crossed unicorn status with a round led by a Value Partners Group private-equity fund. Funds direct Pharma.AI platform plus a 30-asset pipeline; lead candidate Rentosertib (ISM001-055) for idiopathic pulmonary fibrosis remains in clinical studies. via PR Newswire
Who shipped
Cohere kept its enterprise lane crisp — Command A is explicitly tuned for two-GPU deployments at scale. Sesame open-sourced its voice stack — a rare full release after the Maya/Miles demo virality. Insilico Medicine kept the AI-bio crossover funding cycle humming.
Open-source pulse
CSM-1B under Apache 2.0 is the day’s clean open-weights drop — voice models with viral-quality output are rarer than text peers, and the Llama-derived architecture gives easy integration paths. Cohere itself remains closed-weights; the Hugging Face mirror is API-gated.
Money, infra & hardware
Insilico’s $110M Series E was the day’s headline raise. Cohere’s two-GPU pitch is itself an infra story — pushing back against the assumption that frontier-tier capability requires hundred-GPU deployments.
Quiet corners
No major OpenAI, Anthropic, Google, Meta or xAI announcements specifically dated March 13. The frontier-lab attention was still on the OpenAI Responses API and Gemma 3 from earlier in the week.
By the numbers
- 111B — Command A parameters; 256K context
- 2 — H100/A100 GPUs needed to run Command A
- 156 tok/s — peak Command A throughput
- 1M+ hours — Sesame CSM-1B training audio
- $110M — Insilico Series E; unicorn status reached
- Most-mentioned company: Cohere
Compiled by AI Feed’s editor from verified web sources for 13 March 2025.