deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Qwen4-Max reaches public API with 1M-token context as llama.cpp hits 1.0.0

Alibaba moved Qwen4-Max into its public API with a one-million-token context window, while llama.cpp reached 1.0.0 with breaking C API changes, a dev.to roundup reports.

Qwen4-Max reaches public API with 1M-token context as llama.cpp hits 1.0.0

Qwen4-Max reaches the standard API

Alibaba has moved Qwen4-Max out of preview and into general availability on its API, with ordinary token-based pricing, according to a roundup on dev.to. The model is a Mixture-of-Experts system with a reported 2.4 trillion total parameters, of which around 71 billion are active on each forward pass, and the endpoint offers a native context window of one million tokens.

Pricing sits at $1.50 per million input tokens and $4.50 per million output tokens, the report says. For teams that want to push entire document collections through a single request, that cost profile is workable, though applications that generate long completions will feel the output rate quickly.

The change matters mostly for access: developers can now wire the model into production systems directly instead of relying on limited preview access or wrapper services.

llama.cpp stabilises at 1.0.0

The local inference runtime llama.cpp has shipped version 1.0.0. According to the dev.to report, the milestone establishes a settled base for building GGML-compatible model support into native applications. It is not friction-free, though: several breaking changes to the C API mean maintainers of language bindings and downstream tools will need to revise their integration headers and rebuild against the new release.

Projects with existing compilation pipelines should budget for a one-off migration. The roundup's own suggestion is to update headers first and get the build clean before experimenting further.

DeepSeek builds native tools into its endpoint

DeepSeek also expanded its endpoint this week, adding built-in tools for web search that returns citations, search invoked during reasoning steps, PDF parsing and URL reading, exposed through interfaces such as deepseek_web_search and deepseek_read_url. The company additionally published an MIT-licensed static web client that talks to the API straight from the browser, which makes it possible to run a custom retrieval interface without an intermediate Node.js or Python service.

An 80-billion-parameter model on a phone

On the opposite end of the hardware spectrum, the open-source smolmaxxing project ran an 80-billion-parameter MoE model on a OnePlus Nord CE 2 handset using a custom llama.cpp patch that offloads MoE layers. The setup quantises experts to 2 bits and routes layers to control memory, achieving 2.6 to 3.1 tokens per second with peak memory use of roughly 5 GB, per the roundup. That is far below comfortable interactive speed, but it shows phone-class hardware can take part in on-device agent workloads as long as memory traffic is managed carefully.

The same roundup highlights Stageshow, an open-source skill that lets AI coding agents generate React product UI demos — dashboard layouts, mobile screens and animated showcases with scripted cursor motion — directly from natural-language prompts.

Unverified reports to treat with caution

Two items in the roundup arrive with explicit uncertainty attached. Community reports suggest the Codex team is working on internal fixes for recurring capacity errors at busy times, but no official incident report has appeared. Early design leaks also point to an overhauled settings page for Comic artifacts within Gemini Notebook, with no confirmation from Google or outside testers on timing. Neither claim should be treated as settled.

Why it matters

The two headline releases sit at opposite ends of the AI deployment spectrum, and both moved toward practicality. Qwen4-Max places a very large, very long-context model behind standard API calls at rates that make multi-document work affordable, removing the preview-tier hurdle for agents and analysis pipelines. llama.cpp's 1.0.0 tag turns the de facto standard for local inference into a versioned, stable contract — a worthwhile trade for anyone maintaining native integrations, since the breaking changes are a one-time cost.

Combined with efforts like smolmaxxing, the week suggests the gap between managed frontier endpoints and edge hardware is narrowing, even if latency on constrained devices remains the binding constraint. One caution carries across both worlds: long completions cost real money, and unwatched token budgets can grow faster than expected.

  • #qwen
  • #llama-cpp
  • #api
  • #local-inference
  • #open-source

Related posts