New: DeepSeek V4.1 Flash. The architecture built for agents, and the 890 bytes that change an agent bill. Explore the piece →

// changelog

Every shipped
change.

A running log of what we ship, new models, API surface, performance and platform. No marketing, just the diffs.

platform
New

GLM 5.3 replaces GLM 5.2 as the per-key add-on

Same base model, seven points higher on the Artificial Analysis index, where it ties Kimi K3 at the top of the open-weight board. The price and the per-key limits do not move, and the served context goes up: the whole 1M per request instead of 500K. It is text only, so the vision cases stay on Qwen3.6 and Gemma 4.

  • Still €150 a month per API key
  • 1M of context per request, up from 500K
  • Same monthly allowance by tier: 2B, 6.5B and 18B tokens
  • Z.ai’s own GLM-5.3 licence, not MIT
platform
New

GLM 5.2, the strongest open model on the platform

The 744B MIT model now runs on our EU hardware with zero logs. It is an add-on you attach to any enterprise plan rather than something a plan covers: add as many API keys with GLM 5.2 as you need.

  • €150 a month per API key
  • 500K of context, 100 rpm, 3M tokens per minute
  • Monthly allowance by tier: 2B, 6.5B and 18B tokens
platform
New

RAG + Chat

Upload files or connect a repository as a knowledge base, then ask questions about it straight from Chat. Retrieval and model run on the same private infrastructure.

  • Files and GitHub repositories
  • Guide at /docs/knowledge-base
api
NewAPI

Image generation with Flux-2-klein

Generate images from the Helmcode dashboard or from the API, with the same key you already use.

platform
Improved

2× throughput on Qwen 3.6

Speculative decoding is now on by default for qwen3.6, roughly double the tokens per second at the same latency, no change on your side.

api
NewAPI

Reranking endpoint

A dedicated /v1/rerank endpoint for cross-lingual semantic reranking, the missing middle step of RAG (embedding → rerank → LLM).

  • Powered by Qwen3-Reranker-8B
  • 100+ languages
console
Security

Zero-logs attestation in the console

Every API key now shows a live attestation that no prompt or completion content is stored, something your compliance team can screenshot.

platform
New

Dedicated GPU plans

Exclusive NVIDIA B200 hardware inside Helmcode's EU infrastructure, guaranteed throughput, full network isolation and custom models.

  • Custom models & fine-tuning
  • Custom SLA
platform
Improved

Faster cold starts, lower p95

Reworked model loading and routing in the control plane. Cold starts are noticeably quicker and p95 latency is down across the board.

api
NewAPI

Speech: TTS and STT

Kokoro text-to-speech (sub-second latency, 67 voices) and Whisper Large v3 speech-to-text (99+ languages), both behind the same key.

  • /v1/audio/speech and /v1/audio/transcriptions
api
Fixed

Streaming with tool calls

Fixed an edge case where streamed responses could truncate when a tool call and content were interleaved. Streaming is solid across all chat models.

That's everything so far, updated as we ship.