Welcome to GigaElixir Gazette, your 5-minute digest of Elixir ecosystem news that actually matters.
This week in Elixir: ElixirConf EU 2026 wrapped in Malaga with type inference merged into v1.20.0-rc.5, a head-to-head benchmark shows Phoenix LiveView connecting WebSockets 14 times faster than Rails Hotwire, and Nx 0.12 arrives with native C and CUDA extension support.
Also: OTP 29.0 drops with post-quantum SSL key exchange, the Code BEAM Europe 2026 CFP deadline is May 31, and the Pro Tips section covers how logits-level grammar constraints make LLM output validation unnecessary.
. WEEKLY PICKS .
🔥 ElixirConf EU 2026: Type Inference, DurableServer, and 4x Faster Compilation
Malaga surfaced three shifts that matter. Elixir 1.20 type inference arrives without annotations - the compiler understands function signatures, guard conditions, and map key domains, emitting warnings when operations reference keys proven absent. DurableServer decouples process state from node lifecycle using pluggable storage backends. Lazy module loading and parallel dependency resolution deliver up to 4x compilation speed gains. AI orchestration patterns rounded out the track with Legion for code execution by agents and Arcana for retrieval-augmented generation.
⚡ Phoenix LiveView Handles 5x HTTP Load and Connects 14x Faster Than Hotwire
The same collaborative to-do board, built twice. Phoenix LiveView's median response time was 130ms versus Rails Hotwire's 1,484ms - 11 times faster at p50. WebSocket connection time was 30ms versus 418ms - 14 times faster. Phoenix created around 3,200 todos in 70 seconds; Rails managed 640. Memory at 500 concurrent connections was roughly equal at around 70MB each. Phoenix required zero lines of application JavaScript. Rails required around 200 lines across three Stimulus controllers.
🚀 Nx 0.12 Ships C and CUDA Extension Support Through CustomCall Protocol
Nx, EXLA, and Torchx 0.12 are published. The headline addition is the Nx.block abstraction paired with the EXLA.CustomCall protocol - together they let developers provide native C or CUDA implementations for specific sections of Nx computation graphs. EMLX 0.3 (the MLX backend for Nx) ships simultaneously with the updated concurrency model, fused operations via EMLX. Fast, quantized type support, and a zero-copy implementation of Nx.to_binary/1.
📊 OTP 29.0 Drops: Post-Quantum SSL, Secure SSH by Default, Erlang Doctests
Episode 304 covers a busy week beyond the type system. OTP 29.0 is a major release with three notable additions: secure-by-default SSH configuration, post-quantum SSL key exchange algorithms, and Erlang doctests. String processing gets 1.5-5x speed gains through SWAR (SIMD Within A Register) optimizations across Base and String modules. One more item: a dev.exs config tweak that preserves LiveView socket state across hot reloads - no more losing form state mid-development.
🎯 Code BEAM Europe 2026 CFP Closes May 31 - Haarlem, Netherlands
The conference runs October 21-22 in Haarlem, Netherlands. This year's tracks cover data sovereignty, distributed systems, open source sustainability, resilient IoT backends, production realities, and resource efficiency. Proposals run 20 or 40 minutes. First-time speakers get mentorship covering idea evaluation, abstract refinement, and slide structure. The organizers explicitly support candidates from underrepresented groups. Deadline is May 31 - if you have a talk in you, this is the conference to give it at.
💡 Pro Tip
LLM Output Validation Is the Wrong Layer for Schema Enforcement
The standard approach to structured LLM output: generate, parse, validate, retry on failure. It works until the pipeline delivers invalid JSON to a downstream service at an inconvenient time.
Post-generation validation runs after the model has already committed to tokens that violate your schema. Each retry costs tokens and adds latency. And still produces a different violation often enough to matter.
The alternative has existed since 2016 but remained in research papers: logits-level grammar constraints.
Before the model selects each next token, you prune every candidate that would violate a grammar you supply. The model cannot emit output that breaks your schema. No retries, no patch-ups.
Bumblebee and Nx make the approach accessible today. Supply a grammar at decoding time, hook into the logit processing pipeline, and zero out invalid token scores before sampling.
Chris Beck's Code BEAM Europe 2025 talk demonstrates the wiring with working Elixir code, covering three critical use cases: emitting strict JSON configs, generating test fixture data, and producing valid Erlang term syntax, tasks where unconstrained decoding fails often enough to justify the constraint.
BEAM's supervision model completes the picture. Inference processes restart automatically on failure, and grammar constraint pipelines compose cleanly with existing OTP supervision trees.
Remember, for LLM schema enforcement on the BEAM:
Constrain at decode time, not parse time - Post-generation validation lets the model commit to invalid tokens; logits pruning prevents those tokens from being selected
Bumblebee and Nx expose the hook today - The logit processing pipeline is accessible in Bumblebee; no custom runtime or external dependency required
Small but critical tasks benefit most - JSON configs, test data, and Erlang term generation are exactly where unconstrained decoding fails often enough to justify grammar constraints
OTP supervision handles failure - Inference processes under a supervisor restart automatically; grammar constraint pipelines compose with standard OTP supervision trees
. TIRED OF DEVOPS HEADACHES? .
Deploy your next Elixir app hassle-free with Gigalixir and focus more on coding, less on ops.
We're specifically designed to support all the features that make Elixir special, so you can keep building amazing things without becoming a DevOps expert.
See you next week,
Michael
P.S. Forward this to a friend who loves Elixir as much as you do 💜
