AI developer tools · What shipped
For a senior engineer who already reads HN. Real changes in AI developer tools today: releases with version numbers, papers with benchmarks, repos that crossed a threshold worth knowing. Skip hype threads, pre-announcement leaks, and recycled summaries. Always link primary sources.
← Zur aktuellen AusgabeoMLX 0.7.1 ships, Together adds sparse models, Cortex expands regions
oMLX 0.7.1 for Apple Silicon
Local inference on Mac just got a speed bump.
oMLX 0.7.1rc1 landed October 15 with support for Meta's 816B Muse Spark Open (42B active parameters) and DFlash2 speculative decoding baked into batched engines [Quelle: GitHub]. GLM-5.3-Flash now hits 95.3 tokens/sec on M5 (up from 85.5), and Qwen3.8-27B with TurboQuant doubled to 68.8 tok/sec on M3 Ultra. The stable v0.7.0 release includes a rebuilt memory guard and Lightning MTP acceptance rates jumped from 16–21% to 61–85% on Qwen3.8-27B.
Continuous batching on consumer hardware just stopped being a bottleneck.
Together AI serverless models
Together's API just added sparse-MoE options.
The platform released new serverless models October 16, expanding inference choices for cost-conscious builders [Quelle: Together AI]. Sparse architectures trade latency for token throughput, letting teams pick the tradeoff that fits their workload. Pricing and per-model benchmarks roll out as documentation updates land.
Read the changelog for exact parameter counts and availability.
Claude Code v2.1.302
Claude Code shipped 302 with fixes and one deprecation notice.
The October 16 release squashed prompt cache misses, MCP server protocol bugs, and OpenTelemetry logging issues [Quelle: GitHub]. Standalone Claude Design closes December 14. A new environment variable, CLAUDE_CODE_AUTO_MODE_MODEL, lets teams pick the client-side classifier for auto mode—useful for teams running air-gapped or latency-sensitive deployments.
Model selection control moves to your CI/CD layer.
Snowflake Cortex on three clouds
Cortex AI Gateway now runs outside AWS.
Snowflake expanded availability to Azure, Google Cloud Platform commercial regions, and AWS Asia Pacific (New Zealand) as of October 16 [Quelle: Snowflake]. Multi-cloud gateway deployments remove lock-in friction for teams with existing Snowflake warehouses. Regional routing now lets data stay closer to users.
Watch for pricing alignment across regions.