Skip to content

Inkling 2 Small — integration report

What fast-agent verified when adding Inkling 2 Small. The full onboarding record and its evidence are in the fast-agent-model-lab repository.

Model card

name: Inkling 2 Small
vendor: Thinking Machines
id: thinkingmachines/Inkling-2-Small
route: hf.thinkingmachines/Inkling-2-Small:deepinfra
class: flagship
source: https://huggingface.co/thinkingmachines/Inkling-2-Small

Wire contract

Tested 2026-11-14.

Check Result Finding
Each effort accepted, with observable effect; default and off behaviour ⚠️ caveat 66 valid-setting runs correct; adjacent efforts overlap, none remains soft, historical default outlier reproduces at 4x allowance
Reasoning field, streaming shape and replay requirement ✅ pass reasoning_content identified; shared replay succeeds with reasoning, reasoning_content and omission
Default and maximum output tokens; truncation finish reason ⚠️ caveat Omission generated 60014 tokens; both limit fields truncate at 64; separate output ceiling unknown
Enforced context window and request size limits ⚠️ caveat Input rejection confirms 1048576 window; recognized 4.6 MiB reaches Together, 5/6 MiB get router 413
Chunk shape, finish reasons, terminal usage ✅ pass Standard finishes and DONE; terminal usage with include_usage, absent without it
Single, parallel and forced tool calls; streamed arguments ✅ pass Complete streamed get_weather matrix including forced none and real-call continuation; fragment shapes recorded
JSON schema, JSON object, and schema with tools ⚠️ caveat City schema validates via legacy envelope; standard schema rejected and legacy auto-tool cases skip the tool
Each documented modality; unsupported types rejected cleanly ✅ pass Required image fixture passes inline, HTTPS and tool role; sample PDF, MP4 and audio cleanly rejected
Accepted and rejected sampling parameters ✅ pass All nine parameters sent individually; ranges and recommendations recorded including null unknowns
Usage fields and cached-token reporting ✅ pass Usage fields recorded; corrected roughly 5K prefix with different final lines reports 4864 cached tokens on repeat
Proposed model parameters and route profile ✅ pass Evidence-to-field table complete; product recommendations provisional under D1-D4/D7-D8

Wire contract — hf.thinkingmachines/Inkling-2-Small:deepinfra

Tested 2026-11-14.

Check Result Finding
Each effort accepted, with observable effect; default and off behaviour ⚠️ caveat Seven efforts and omission accepted, 3/3 correct each; none empties the reasoning channel and writes the working into visible content; bogus rejected 422
Reasoning field, streaming shape and replay requirement ✅ pass reasoning_content in stream and message; replay as reasoning, reasoning_content and omitted all answered
Default and maximum output tokens; truncation finish reason ⚠️ caveat Omission generated 60593 tokens (stop); both limit fields truncate at 64; max_tokens up to 2000000 accepted, ceiling not discoverable
Enforced context window and request size limits ⚠️ caveat 1048576 window enforced; 287050-token prompt answered; router body cap exactly 5242880 bytes, configured 4718592 stays
Chunk shape, finish reasons, terminal usage ✅ pass stop/length/tool_calls, DONE, usage in its own chunk only with include_usage
Single, parallel and forced tool calls; streamed arguments ⚠️ caveat Single, parallel, forced, forced at none and continuation correct on the wire; explicit-null placeholder fragments break the SDK accumulator, so the route needs manual stream mode
JSON schema, JSON object, and schema with tools ⚠️ caveat Standard json_schema valid 8/8 incl. non-strict, streamed and max; legacy envelope unreliable; schema with tools skips the tool 4/4
Each documented modality; unsupported types rejected cleanly ✅ pass PNG/JPEG/WebP and HTTPS images correct; PDF answered correctly (route-only); video/audio 400; image in tool message 422
Accepted and rejected sampling parameters ✅ pass All nine parameters accepted; stop honoured
Usage fields and cached-token reporting ✅ pass Usage plus estimated_cost; cold 0 then 6144 cached tokens on a fresh ~6K prefix
Proposed model parameters and route profile ✅ pass Together profile without the legacy schema envelope, plus route-level manual stream mode; un-pinned router route provisionally gets the DeepInfra profile

Harness conformance

Tested 2026-11-14.

Check Result Finding
Streamed text turn ✅ pass 1 pass
Every ladder value through the alias; effective wire value recorded ⚠️ caveat 7 pass, 1 warn
Tool call, execution, continuation, final answer ✅ pass 3 pass
History and reasoning replay across turns ✅ pass 1 pass
Structured output validated against the schema ✅ pass 2 pass
Each supported modality understood; unsupported types handled ✅ pass 3 pass
Usage and context accounting ✅ pass 1 pass
Raw request and stream traces inspected ✅ pass Requests/chunks inspected; route, reasoning order/replay, tool IDs/fragments, finishes and usage verified; raw traces deleted
Picker, effort cycling and check commands ✅ pass Picker, all seven efforts, off, invalid-value rejection and full F6 wrap verified; seven distinct renderer styles; 106 unit tests pass

Harness conformance — hf.thinkingmachines/Inkling-2-Small:deepinfra

Tested 2026-11-14.

Check Result Finding
Streamed text turn ✅ pass 1 pass
Every ladder value through the alias; effective wire value recorded ⚠️ caveat 7 pass, 1 warn
Tool call, execution, continuation, final answer ✅ pass 3 pass
History and reasoning replay across turns ✅ pass 1 pass
Structured output validated against the schema ✅ pass 2 pass
Each supported modality understood; unsupported types handled ⚠️ caveat Suite 3 pass; manual: PDF correctly removed, but the follow-up turn returned empty stops in 6/8 DeepInfra attempts (2/4 runs failed), 0/3 on Together
Usage and context accounting ✅ pass 1 pass
Raw request and stream traces inspected ✅ pass Requests/chunks inspected; wire fields, reasoning order and replay, tool fragments, finishes and usage verified; raw traces deleted
Picker, effort cycling and check commands ⚠️ caveat Static checks only (task forbids starting agents); no picker entry by design; ladder and F6 are model-level and verified on Together

Structured output — Wire contract

Evidence: fast-agent-model-lab/onboarding/models/inkling-2-small/evidence/contract-structured.json.

The prescribed schema requires city: string, population_millions: number, and disallows extra properties. Prompt: Give the largest city in Japan and its population in millions. Complete visible content was parsed and checked using jsonschema.validate, not just HTTP status or excerpts.

Mode (two repeats each) Status Valid Tool first?
Standard json_schema, strict 400 / 400 false / false n/a
json_object, schema in prompt 200 / 200 true / true n/a
Legacy json_object + schema 200 / 200 true / true n/a
Standard schema + get_weather 400 / 400 false / false no response
Standard schema at max effort 400 / 400 false / false n/a
Legacy schema + get_weather 200 / 200 true / true no / no
Legacy schema at max effort 200 / 200 true / true n/a

Standard envelope errors report missing required response_format.json_schema; legacy succeeds. The prior conflicting enum/prompt probes still support actual schema enforcement beyond accidental compliance. Warn: legacy auto tools are skipped despite the request to use get_weather first. D2 remains legacy schema + defer, provisional; later harness conformance must prove tool-then-final-schema behavior. This stage does not establish strict-envelope equivalence or population factuality.

Structured output — Wire contract: deepinfra

Evidence: fast-agent-model-lab/onboarding/models/inkling-2-small/evidence/contract-deepinfra-structured.json.

Mode Valid Tool called
json_schema strict 2/2 –
json_schema non-strict 2/2 –
json_schema strict, streamed 2/2 –
json_schema at max 2/2 –
json_object + schema in prompt 2/2 –
legacy {"type": "json_object", "schema": …} 1/2 (plus 0/1 in the comparison) –
json_schema + get_weather 2/2 0/2
json_schema + get_weather, streamed 2/2 0/2

The standard nested envelope is the right form here; the legacy envelope is accepted but does not enforce the schema (one answer was prose). With a schema and an auto tool the model answers from memory without calling the tool, as on Together: keep structured_tool_policy="defer". Graded warn per the contract playbook because of the skipped tool; the harness defer policy is assessed separately in conformance.