Coding models and inference economics #69

Today's Letter

1. Mistral AI, Leanstral 1.5 proof model release
2. Meta, Muse Spark update targets coding and agents
3. Wafer reports MI355X cost edge in GLM-5.2 inference

Mistral AI, Leanstral 1.5 proof model release

• Mistral AI published Leanstral 1.5 on July 2, 2026 as an Apache-2.0 licensed model for formal proof generation.
• The primary announcement describes the model as having 119B total parameters and 6B active parameters.
• Mistral says Leanstral 1.5 reaches saturation on miniF2F and solves 587 of 672 PutnamBench problems.
• The release targets Lean-based formal verification workflows and focuses on broader access to proof-oriented models.
• Hugging Face is listed in related release materials as a distribution path for the model.
• The main disputed field is active parameter count, but the body follows the primary source figure of 6B active parameters.

Source: mistral.ai
More: news.hada.io · testingcatalog.com · the-decoder.com

Meta, Muse Spark update targets coding and agents

• Meta said the next Muse Spark update will improve coding and agentic AI capabilities.
• Chief AI Officer Alexandr Wang posted that the release is coming soon and is meant to compete more directly with leading models.
• The upcoming version is codenamed Watermelon and uses substantially more compute than the previous Muse Spark release.
• Business Insider, citing anonymous sources, reported Wang said Watermelon had caught up with OpenAI's GPT 5.5 in internal discussion.
• Meta plans to roll out the model through Meta AI and a new API.
• Analysts said a stronger Meta model could give enterprises another option alongside OpenAI and Anthropic.
• Enterprise adoption still depends on real-world coding quality, agent reliability, security, governance, and developer ecosystem maturity.

Source: infoworld.com
More: ababnews.com · americanbazaaronline.com

Wafer reports MI355X cost edge in GLM-5.2 inference

• According to the report, Wafer benchmarked GLM-5.2 on AMD MI355X and said the setup delivered lower cost per node than NVIDIA Blackwell-class systems.
• The report says MI355X is about 2.75x cheaper per GPU than B300, while reaching 2626 tok/s/node and 2.4 rps on a 20k input, 1k output workload with a 60% cache hit rate.
• In single-stream testing, Wafer reported 213 tok/s with a 10k input and 1.5k output profile, described as below the top of the Artificial Analysis leaderboard but stronger on performance per dollar.
• The write-up attributes most of the gain to MXFP4 quantization, speculative decoding fixes, and MoE kernel tuning rather than custom kernels.
• Wafer also argues that AMD's main constraint is now software support and day-0 model readiness, not raw silicon competitiveness against the Instinct MI350 and Blackwell generation.
• The figures were presented as single-source benchmark results from a TensorWave MI355X deployment and were not cross-verified by secondary reporting in the provided materials.

Source: news.hada.io

Jocoletter curates AI, software, and product trends for developers and builders.

#Meta #MistralAI #Wafer

Subscribe to Jocoletter