LABEL CHECK / DOSSIER

Ollama

https://github.com/ollama/ollama·verified at v0.30.11 (2026-06-25)·last verified 2026-06-28·volatile (re-verified per release (≤2 weeks))

11claims sampled

6as labeled

4partial

1not as labeled

0not implemented

6open deltas

0verified closed

1closed partial

closure trend: context defaults moved materially across releases (vendor-side, partially closing OL-02); tool-calling deltas show no TD-re-verified closures

Verdict

Ollama's simplicity, model-library, platform, and context-default documentation claims hold as labeled, but its tool-calling label is contradicted where it matters for agents — documented streaming tool support drops tool_calls chunks silently in production, and the get-running-in-minutes agent framing inherits a default context below the vendor's own agentic guidance on common hardware.

Claim ledger — 11 rows

#statusclaimedobservedevidenceverified

1as labeledThe easiest way to build with open models Full detail →Source-reviewed: dominant local LLM runtime; every major local-first agent framework in the corpus (Goose, Continue, AnythingLLM) treats Ollama as the default inference backend; onboarding is a single `ollama` command… Full detail →[1][2]—

2partialGet up and running with OpenClaw, Claude Code, and more in minutes using open models Full detail →Release-verified: `ollama launch` bootstraps full agent toolchains in minutes (Hermes since v0.21.0, Cline CLI auto-install since v0.30.2, Hermes Desktop since v0.30.7). Full detail →[1][2][3]—

3partialOpenAI compatibility: /v1/chat/completions with chat completions, streaming, JSON mode, vision, and tools Full detail →Docs-reviewed: the endpoint inventory is implemented, and the docs themselves scope out tool-choice parameters and stateful Responses. Full detail →[1][2][3]—

4not as labeledTool calling with streaming (documented as a supported pattern with Python and JavaScript examples)Full detail →Contradicted by production evidence: when streaming is enabled (the default), Ollama returns an empty content chunk with finish_reason "stop" instead of tool_calls delta chunks — every tool-dependent agent skill (web… Full detail →[1][2][3][4][5]2026-07-12

5partialTool calling as a first-class capability (single-shot handling "only recommended for models which only return a single tool call")Full detail →Holds for capable models in non-streaming mode: native GBNF grammar enforcement since v0.5; Qwen3-14B reaches F1=0.971 tool selection (source-reviewed corpus benchmark). Full detail →[1][2]—

6as labeledDefault context length based on VRAM — "< 24 GiB VRAM: 4k context, 24-48 GiB VRAM: 32k context, >= 48 GiB VRAM: 256k context"; agent and coding tasks "should be set to at least 64000 tokens"Full detail →As documented — source-confirmed at HEAD 2026-06-11: envconfig/config.go defines OLLAMA_CONTEXT_LENGTH defaulting to 0 with VRAM-based resolution described as "4k/32k/256k based on VRAM"; release-verified that v0.15.5… Full detail →[1][2][3]2026-07-12

7as labeledRun entirely offline for mission critical work Full detail →Holds for local models: local inference requires no external calls, and the Hermes agent integration (`ollama launch hermes`, v0.21.0) is documented as running fully offline with no API keys — release-verified. Full detail →[1][2]—

8partialYour data stays yours / Your data is never trained on Full detail →Structural for local models. Materially qualified since v0.18.0 (2026-03-14, release-verified): "Ollama's cloud models no longer require downloading via `ollama pull`. Full detail →[1][2]2026-07-12

9as labeledModel library: "Run and chat with Gemma 4"; "See ollama.com/library for the full list."Full detail →Docs- and release-reviewed: the library carries the major open model families and stays current within days of upstream releases — Gemma 4, the Qwen3 family (including sub-14B tool-capable variants referenced in the… Full detail →[1][2]—

10as labeledPlatform support: macOS, Windows, Linux, and an official Docker image (ollama/ollama on Docker Hub)Full detail →Release-verified active maintenance on every claimed platform: ROCm 7.2.1 update for Linux/AMD (v0.20.8), native Windows configuration-path fixes in the v0.30.x line, Apple Silicon backend actively developed, official… Full detail →[1][2]—

11as labeledOllama has a REST API for running and managing models.Full detail →Confirmed: the REST API is the integration surface for the framework ecosystem, and source-reviewed corpus evidence shows API compatibility was maintained across the Apple Silicon MLX backend switch (March 2026) —… Full detail →[1][2]—

Delta ledger — 6 open

OL-01opened 2026-06-11

Default context on <24 GiB VRAM hardware (4,096 tokens) is below the vendor's own >=64,000 agentic guidance; truncation remains silent

closes when:Default context on <24 GiB VRAM hardware meets the vendor's own agentic guidance (>=64,000 tokens), or context truncation surfaces an explicit warning or error to the caller — TD re-verified on a current release

evidence:[1][2][3]UPDATE 2026-06-28: v0.30.9 errors on single-message context overflow (partial progress); multi-turn accumulation truncation still silent — delta remains openRe-verified open 2026-06-28 against v0.30.11

OL-03opened 2026-04

Streaming layer returns empty content with finish_reason "stop" in place of tool_calls delta chunks — all tool-dependent agent skills fail silently

Re-checked 2026-06-11: vendor docs now document tool calling with streaming, and the closest upstream issue (#16212) was closed "completed" the same day it was filed after its reporter marked it opened in error. Full detail →

closes when:An agent receives tool_calls delta chunks (correct incrementing indices, finish_reason "tool_calls") over a streaming response in default configuration, with no stream:false workaround — TD re-verified on a current release

evidence:[1][2][3][4][5]Re-verified open 2026-06-28 against v0.30.11

OL-04opened 2026-03-15

Above ~5-6 active tools, qwen3-coder switches from JSON to XML tool-call format and some models stop attempting tool calls — silent degradation, no error signal

closes when:qwen3-coder emits parseable JSON tool calls (or Ollama's parser transparently normalizes the XML form) with more than 6 active tools, TD re-verified — or the runtime surfaces an explicit signal when a model abandons tool-call formatting

evidence:[1]Re-verified open 2026-06-28 against v0.30.11

OL-05opened 2026-03-16

`:cloud` tag references route to cloud endpoints without a prior `ollama pull` (v0.18.0) — local-only deployments can route data externally silently

closes when:Referencing a cloud-tagged model from a configuration that has never pulled it requires an explicit opt-in, or fails loudly before any external request is made — TD re-verified

evidence:[1]Re-verified open 2026-06-28 against v0.30.11

OL-06opened 2026-04-21

Chat template bugs corrupt tool schemas passed via /api/chat before inference — malformed tool definitions upstream of token sampling, unaffected by GBNF enforcement

closes when:Tool schemas pass through /api/chat chat templates intact for the affected models — TD re-verified on a current release

evidence:[1]Re-verified open 2026-06-28 against v0.30.11

OL-07opened 2026-06-26

Tool parameter literally named "name" is silently dropped (empty content, finish_reason stop) — Mistral-format models, v0.30.10 (#16932)

closes when:A tool whose function schema includes a parameter named 'name' is invoked normally (tool_calls returned) rather than silently dropped, TD re-verified on a current release

evidence:[1]

Closed deltas — 1 (falsification-gated)

OL-02closed partial

Flat low context default (2,048, later 4,096) silently breaking multi-turn tool calling

closed by:v0.6.7 (2025-04-26) raised the default to 4,096; v0.15.5 (2026-02-03) replaced flat defaults with VRAM tiers (4k/32k/256k) — release-reviewed and source-confirmed at HEAD 2026-06-11

Methodology

sampling: Label-check methodology (Method §2, D0086) — material-claims rule: a claim is sampled if it appears on the product's own label and would plausibly influence an adoption decision. Sampling frame: ollama.com homepage, the ollama/ollama README, docs.ollama.com capability pages (tool calling, OpenAI compatibility, context length), and release notes v0.6.7-v0.30.7, captured 2026-06-11. The audited product is the local runtime; cloud subscription capacity and pricing claims (Pro/Max tiers) are out of frame. 11 material claims found; all sampled.
verification depth: source-reviewed — not executed by Theory Delta
verification statement: Evidence is drawn from Ollama's source at HEAD (envconfig/config.go), official docs, release notes, the issue tracker, and a third-party production report; context-length defaults and the streaming-bug issue state were re-checked against source and docs on 2026-06-11 and again on 2026-06-28 (v0.30.11). Theory Delta has not executed Ollama's tool-calling paths in its own environment for this dossier. Star counts (111K+ reported 2026-05-26) are treated as unreliable and excluded from assessment. Decay class: volatile (daily release cadence observed).
backing confidence: empirical
strongest case against: Ollama ships daily; the streaming tool_calls failure rests on one third-party production report whose closest upstream issue was retracted by its own reporter, and the vendor's docs now document streaming tool calling — TD re-verification could move claim 4 to partial or close OL-03. The context-default delta rests on the vendor's published VRAM tiers; the actual hardware-detection behavior was not executed. The MLX backend switch (March 2026) may have changed streaming behavior on Apple Silicon in either direction.