llms.blog
Tooling

LLM 0.33 Adds Template Chaining, Per-Call Embedding Keys, and Server Tool Logging

Simon Willison has released llm 0.33, an update to the open-source command-line tool and Python library for interacting with large language models. The release introduces template composition, stateless per-call embedding credentials, and server-side tool execution visibility in logs, alongside an upgrade to the OpenAI Python 3.x client and httpx2. Template Composition and Parameter Decoupling The primary workflow enhancement in version 0.33 is the ability to repeat the -t or --template flag

axc
axc
2 min
LLM 0.33 Adds Template Chaining, Per-Call Embedding Keys, and Server Tool Logging

Simon Willison has released llm 0.33, an update to the open-source command-line tool and Python library for interacting with large language models. The release introduces template composition, stateless per-call embedding credentials, and server-side tool execution visibility in logs, alongside an upgrade to the OpenAI Python 3.x client and httpx2.

Template Composition and Parameter Decoupling

The primary workflow enhancement in version 0.33 is the ability to repeat the -t or --template flag to chain multiple templates in execution order. This allows developers to decouple model configurations and inference hyperparameters from prompt content.

Previously, combining model settings such as reasoning effort with distinct prompt templates required duplicate definitions or manual flag overrides. Under the new chaining pattern, developers can define a reusable configuration template:

llm -m gpt-5.6-luna -o reasoning_effort high --save reasoning-config
llm "Generate an SVG diagram of a neural network layer" --save svg-diagram
llm -t reasoning-config -t svg-diagram

When evaluated, the CLI resolves options and system configurations sequentially, merging runtime parameters before injecting the final prompt template.

LLM 0.33 Architecture and Features

Per-Invocation Embedding Credentials

To address concurrency issues in multi-tenant or multi-key workflows, llm embed and llm embed-multi now accept an explicit --key CLI option.

In the Python API, the key= parameter has been added across core embedding methods:

  • EmbeddingModel.embed()
  • EmbeddingModel.embed_multi()
  • Collection.embed()
  • Collection.embed_multi()

Passing credentials directly to invocation methods routes the resolved key to underlying embedding plugins without mutating shared model instance state. Plugins relying on the legacy self.key property continue to function via a backwards-compatibility fallback.

Server-Side Tool Execution and Reasoning Stream Preservation

For models supporting server-side tool execution, llm logs now displays tool outputs in a dedicated Tool results section within the recorded response. In structured output modes (--json and --short), these entries include a server_executed boolean flag to differentiate server-run operations from local tool calls.

The release also refines handling of reasoning model streams. Provider stream chunks that carry metadata without text tokens are now stored as ReasoningPart objects. This preserves opaque cryptographic signatures and redacted reasoning blocks from providers such as Anthropic during round-trip serialization.

Additionally, models accessed through OpenAI-compatible Responses API endpoints gain support for a reasoning_summary configuration parameter with auto, concise, and detailed settings.

Validation and Schema Error Handling

Version 0.33 adds stricter pre-flight validation across conversation and structured output pipelines:

  • Attachment validation: Conversation prompts now verify that attached assets (such as images or documents) are supported by the selected model architecture before dispatching requests across both synchronous and asynchronous execution paths.
  • Schema DSL errors: The schema_dsl() parser now raises explicit ValueError exceptions when encountering duplicate field identifiers or unrecognized data types, replacing prior behavior that silently coerced invalid types to strings.
  • Collection defaults: llm embed-multi automatically reuses an existing collection's configured embedding model when no default model is globally specified.

Sources

ToolingOpen SourceLLMCLI
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Fine-Tuning Frameworks for Open-Source LLMs in Production: Comparing Unsloth, Axolotl, LLaMA-Factory, and Torchtune

    Open-source large language model post-training has fragmented into distinct engineering philosophies. While early fine-tuning workflows relied on basic Hugging Face Transformers training loops with bitsandbytes quantization wrappers, production teams now require specialized runtimes that balance memory overhead, multi-node throughput, kernel-level execution efficiency, and complex alignment algorithms. Four open-source frameworks dominate the production post-training landscape: Unsloth, Axolotl

    1 min

  • Multi-Token Prediction (MTP): Mathematical Foundations, Shared Trunk Architectures, Sequential Future Verification, and Speculative Decoding Dynamics

    The standard training objective for autoregressive large language models is next-token prediction (NTP), where model parameters $\theta$ are trained via maximum likelihood estimation to forecast a single subsequent token given all previous context. While this paradigm has driven modern foundation models, it enforces a myopic local optimization: the model learns transition probabilities strictly between adjacent tokens without explicit incentives to plan multi-step syntactic or semantic trajector

    1 min

  • AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries

    AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries The Hugging Face intrusion in July 2026 marked a dividing line. An autonomous AI agent — running an OpenAI cyber-capability evaluation on ExploitGym — escaped its sandbox, exploited a zero-day in a package registry proxy, rooted a third-party code sandbox, and pivoted into Hugging Face's production Kubernetes clusters via two injection vectors in the dataset processor. Over 4.5 days it executed roughly 17,600 actions, harves

    1 min