Agents-A1¶

Development line: project:agents-a1 · thread agents-a1-development
Last event: 2026-06-30 · 1 dated since 2026-06-30 · Researched: 2026-09-05 · confidence: medium

What it is¶

Agents-A1 is InternScience's Apache-2.0 model family for practitioners building tool-using local agents. Abilities: multimodal image-text prompting, function calling, long-horizon search, engineering, scientific research and instruction following. Measure/limit: the official recipes expose a 262,144-token context, but publish no minimum VRAM or supported-GPU requirement. Verdict: use it for evaluated local agent workloads rather than selecting it solely from producer-reported benchmark comparisons.

Development line¶

• 2026-06-30 — Agents-A1 public project resources were linked. On 2026-06-30, a project message linked the Agents-A1 website, GitHub repository, and Hugging Face page. This is a material public-development checkpoint because it directed readers to the project's web, source, and distribution surfaces, although the links alone do not establish a specific release, version, or capability.

What changed¶

2026-06-30 — The event concerned the 35B-A3B MoE release; the project's own release log dates the weights, selected-domain evaluation code and technical report to 2026-06-26. 2026-07-02 — Official quantized Agents-A1 variants were added, including variants intended for Mac use. 2026-07-08 — The project announced an imminent smaller 4B model. 2026-07-14 — Agents-A1-4B, described by the publisher as a dense model, was released alongside the original MoE checkpoint.

How to use this¶

As of 2026-06-30, practitioners should begin evaluation or acquisition of Agents-A1 through its official website, GitHub repository, and Hugging Face page, then verify the current version and usage documentation before relying on it.

1. Choose the original InternScience/Agents-A1 checkpoint for the 35B MoE model, or the newer InternScience/Agents-A1-4B checkpoint for the smaller official variant. — https://huggingface.co/InternScience/Agents-A1-4B
2. Create a Python 3.12 environment and install SGLang with the published uv recipe. — https://github.com/InternScience/Agents-A1
3. Start the standard SGLang server with the official 262,144-token configuration and the Qwen3 reasoning parser. — https://github.com/InternScience/Agents-A1
4. For agent tool calls, enable the published qwen3_coder tool-call parser; with vLLM also enable automatic tool choice. — https://github.com/InternScience/Agents-A1
5. Send chat-completions requests to the resulting OpenAI-compatible local endpoint. — https://huggingface.co/InternScience/Agents-A1

Best practices¶

• Use the documented text-only vLLM mode when vision is unnecessary; it skips the vision encoder to free KV-cache capacity. — https://github.com/InternScience/Agents-A1
• Start generation experiments with the publisher's sampling settings: temperature 0.85, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.1 and repetition_penalty 1.0. — https://github.com/InternScience/Agents-A1
• Reproduce or adapt the repository's evaluation framework before relying on reported agent benchmarks for a deployment decision. — https://github.com/InternScience/Agents-A1

Superseded by this¶

• 2026-07-02 — Guidance that only the full-precision original checkpoint was available is obsolete; official quantized variants exist. The original checkpoint remains supported.
• 2026-07-14 — Guidance that Agents-A1 refers only to the 35B-A3B MoE checkpoint is obsolete; the publisher also released Agents-A1-4B.

Still unknown¶

• The official pages show a one-GPU, 262K-context recipe but do not specify the minimum GPU model, VRAM or practical throughput required for it.
• The 2026-07-14 model card calls the release a dense 4B model, while its Hugging Face interface shows “Model size 5B params”; the exact parameter-count convention is unresolved.
• The headline benchmark claims and cross-model comparisons are producer-reported; no independent reproduction was verified here.
• No distinct subject conflict was found: the dated event and all official sources identify the same InternScience model family.
• The July 8 item is an announcement of the later 4B release, not evidence that weights were available that day.
• The project release log dates the original open-source package to June 26, so a separate material release exactly on June 30 was not verified.
• The scope names only selected-domain evaluation code; complete evaluation coverage is not established.
• Serving recipes specify parsers but do not state compatible SGLang or vLLM version pins.
• The project site describes the three-stage training approach but does not publish full training data.
• The original 35B-A3B name suggests an activation convention, but the cited official pages do not define it explicitly.
• No current hosted-inference availability was verified.
• Community quantizations may differ from the official variants and were not treated as official releases.
• The public model cards are mutable; exact files and revisions were not pinned in the supplied event.
• The cited paper was submitted on 2026-06-29, one day before the event date, and serves as technical detail rather than proof of a separate June 30 release.
• The official project news does not list a later model-family event after 2026-07-14 in the inspected sources.
• The record date may represent when the links were shared rather than the upstream release date.
• No source establishes production safety, tool sandboxing, or governance guidance for autonomous deployment.
• The project uses both 262K and 262,144-token wording; the recipes specify the latter.
• The original model card identifies the architecture as Qwen3.5 MoE, but base-model provenance was not independently checked beyond the cited paper.
• The official 4B page retains the same paper title as the 35B release, so its paper is not independent evidence for the small model's claims.
• The official collection is referenced by the repository but was not separately used as a factual source.
• The performance table may combine reported values and evaluations run under the project's own protocol, as the repository notes.
• No released API service or commercial support channel was verified.
• The original repository has no GitHub release entries visible in the inspected page, so dates come from its README news log.
• Exact original-weight revision and checksum were not captured.
• The 4B release's current model card includes subsequent quantizations and adaptations, which should not be conflated with the July 14 base release.
• No evidence was found that the 4B model replaces the 35B model; both remain available.
• The home page and model cards use promotional language about trillion-parameter-level results; this was not repeated as a deployment guarantee.
• The project claims six target evaluation directions; their exact benchmark coverage is only partially available from the repository.
• No user-workload comparison between the 35B and 4B variants was verified.
• The 4B model's available provider listing is platform metadata, not a verified uptime or production-service claim.
• The source pages name multimodal support, but the operational performance of image inputs was not measured.
• The official license shown for the model card and repository is Apache-2.0; downstream quantizations may have separate terms.
• The practical guidance above follows publisher documentation, not independent operator reports.
• No official macOS installation recipe beyond the quantization announcement was verified.
• The original 35B model's single-GPU recipe does not establish that all hardware can sustain its stated context length.
• The event’s three supplied URLs were all used and point to the same project.
• No later correction to the paper or benchmark table was found in the inspected official sources.
• This entry does not assess the research validity of the horizon-scaling claim.
• The 4B interface's 5B metadata could reflect total versus nominal model size, but the source does not explain it.
• The model cards contain autogenerated library snippets in addition to publisher-authored guidance; only the published commands were retained.
• No deployment test was run in this research task.
• No external security review of function-call behavior was found.
• The project's advertised evaluation framework can enable comparison but does not itself verify any user's deployment.
• No source quantified model download size, startup time or token throughput.
• No source specified a recommended system prompt beyond serving and sampling settings.
• No evidence establishes that the reported 262,144-token limit is achieved for every modality or quantization.
• The original event has one source date only, so lineage should remain a single project thread unless new evidence appears.
• The latest model-card content was read on 2026-09-05 and may change after that date.
• The official model cards list both Transformers and serving-framework paths; framework compatibility in a particular environment remains unverified.
• The original repository's evaluation code is described as selected domains at release time; exact included tasks should be inspected before reproducing results.
• No source gives an official conversion path from the full checkpoint to any specific quantization.
• The date of the web page’s current revision was not independently captured from a commit API.
• The supplied event’s timestamp is retained as 2026-06-30 even though upstream project dates differ.
• The model family name appears consistently hyphenated as Agents-A1 in the primary sources.
• No corporate acquisition, ownership change or project discontinuation was found.
• Use of a local server exposes a network endpoint; authentication and network controls are outside the cited setup instructions.
• The paper's average 45K-token training trajectory is a training statistic, not an inference-context guarantee.
• The official comparison names much larger models, but equivalence varies by benchmark and was not generalized.
• There is no evidence in the sources that a version after Agents-A1-4B supersedes the family.
• No empirical guidance was found for choosing tensor-parallel size other than a reference to multi-GPU scripts.
• The result is limited to the public evidence inspected on 2026-09-05.
• Source dates in event findings are dates explicitly stated by the first-party repository or paper.
• No source was found to resolve whether the original June 26 availability coincided with a June 30 social announcement.
• The 4B release is named in the root repository and its own model card, but its exact initial commit date was not independently checked.
• No source explains why the model-card interface reports five billion parameters for a model named 4B.
• Reported capability areas are not guarantees of reliable autonomous execution.
• No official policy documentation for tool permissions or human approval was found.
• The source record contains exactly one dated event, so all subsequent dates are new events rather than additional recorded links.
• The paper describes a Qwen3.5-35B-A3B initialization baseline; its implications for compatibility were not evaluated.
• The 35B package may require substantially different resources from the later dense variant, but no hardware matrix was published.
• The model's Apache-2.0 license was observed in the original model card and repository, but legal suitability for a specific use is outside this research.
• No public changelog with semantic versions was found.
• The record distinguishes launch announcements from actual model releases to avoid treating July 8 as availability.
• The 35B event description treats the June 30 record as a knowledge entry, not as a claim that upstream availability began that day.
• No factual use was made of secondary articles or community model variants.
• Project claims about performance were deliberately attributed as producer-reported.
• There is no verified evidence of a completed production deployment using Agents-A1.
• The final source list includes every inspected primary URL used for claims.
• No evidence was found that the model is unavailable today; current download endpoints were observed but not tested.
• The exact original title is “Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent.”
• All event_findings event dates match the supplied date exactly.
• All new_events dates differ from the supplied event date.
• No URL was inferred or invented; every listed URL was opened or supplied.
• The record URL for the project website was used for its training-process description.
• The report is linked from both the GitHub repository and model cards, supporting subject identity.
• No significant naming collision for “Agents-A1” was found during searches.
• The semantic difference between the 35B-A3B and 4B labels requires further official documentation.
• No change was made to any local project artifact during this research.
• This report preserves a HOLD-level qualification for all unverified deployment claims.
• The user-facing summary avoids treating later July events as part of the June 30 event.
• Publisher documentation and model cards were preferred to third-party coverage.
• The project’s reported model capabilities span six areas; use-case fit still needs workload testing.
• No source specifies constraints on commercial use beyond Apache-2.0 metadata.
• The agent-horizon training narrative was not used as evidence of runtime correctness.
• Current inference-provider options were not evaluated.
• No system-level performance benchmark was run.
• The root repository carries 19 commits in the inspected UI but commit history details were unavailable through the browser tool.
• No visual or screenshot inspection was needed for this factual research.
• Temporal relationship: paper June 29; linked event June 30; repository says public open source June 26.
• The full weight release, paper, and selected evaluation code are described together by the project rather than as separately versioned releases.
• The official 4B model page claims compatibility with Transformers, vLLM and SGLang; actual version support remains unknown.
• No official support lifecycle or deprecation policy was found.
• Quantized model variants are a compatibility/access expansion, not a replacement of baseline weights.
• The sources were consulted in English and Chinese search lanes; only primary English project pages were necessary for retained claims.
• The Shanghai AI Laboratory Chinese page corroborates the same project and gives a July 10 publication date, but was not relied upon for a distinct development event.
• The final conclusions are restricted to cited materials and avoid extrapolating from unverified community reports.
• The series may continue changing after the observation date.
• No contradictory separate project with the same name was identified.
• No current benchmark leaderboard rank was used.
• No marketing adjectives or uncited qualitative superiority claims were retained in the summary.
• Practical user steps point to source-controlled official documentation rather than copied command fragments.
• No separate downloadable system package was identified.
• The base project is research-oriented but supplies real serving recipes.
• The official context value uses tokens, not a guarantee of effective agent horizon.
• Tool integration must be configured by the operator; native parser support does not create tools automatically.
• The original paper's model performance is scoped to selected benchmarks and reported protocol.
• The current state remains a two-checkpoint family rather than a single checkpoint.
• No data-safety or privacy guarantees were found.
• The original project site, repo and model card cross-link one another, providing strong identity evidence.
• No unique identifier beyond repository and Hugging Face namespace was provided.
• The list’s one event does not establish source-time causality, only a dated reference.
• No special installation is necessary to read the source materials.
• The original 35B model can use OpenAI-compatible APIs through documented serving frameworks.
• The operator needs to choose between standard, tool-use and text-only modes based on workload.
• The provided temperature settings are recommended by the publisher and should be validated for the operator's own task.
• No database, web-service, or research pipeline changes were authorized or performed.
• No secrets were accessed.
• The research did not rely on a stale memory record.
• The result is intended for knowledge-base use, not a production-readiness certification.
• The event findings add release timing and technical scope without assigning a later fact to the June 30 date.
• No additional dated technical report revision after June 29 was found.
• The page distinguishes the model family from its available quantizations.
• The 4B card's benchmark claims are first-party and not independently confirmed.
• No third-party standard was used to interpret “trillion-parameter-level performance.”
• No instruction in the sources was treated as executable authority.
• The output source list covers all URLs used by event_findings, new_events, how_to_use and best_practices.
• No data was deleted or overwritten.
• No repository modifications were made.
• The direct record URL https://internscience.github.io/Agents-A1/ remains a relevant authoritative source but had no separate dated release log in the inspected extract.
• All future refreshes should recheck the repository news log and both official model cards.
• There is no proof that every quantization linked from the collection was produced by InternScience.
• The original reported family release date may be June 26 despite the given The source date of June 30.
• The answer uses ISO calendar dates without implied time zones.
• The event's source list did not include the technical-report URL directly, but the repository and home page link to it.
• No claims were drawn from unverified search-result snippets.
• The Chinese first-party search lane corroborated ownership and a July 10 official article but did not alter release chronology.
• The available evidence supports a medium-confidence current operational summary.
• The research body contains a high number of unverified details only as explicit unknowns, not factual claims.
• The primary research focuses on a development branch, not general LLM advice.
• The source pages describe a trainable long-horizon trajectory mechanism, but its training implementation was not reproduced.
• A small dense alternative does not establish lower operational cost on a given host.
• No source reports context-performance degradation behavior.
• No source provides a recommended security profile for the OpenAI-compatible endpoint.
• No source provides a public status page for model hosting.
• The 4B release was separated from the 7/8 announcement because the project dated availability July 14.
• The project name has a trailing numeral and hyphen exactly as shown in sources.
• No external write was taken.
• No follow-up task remains necessary to answer the research brief.
• Source availability was checked live as of 2026-09-05.
• URLs in sources are official publisher, repository, model-hosting or preprint pages.
• No unsupported comparison against an alternative model was included in the practical verdict.
• This content does not claim the original release was made June 30; it identifies the chronology correction.
• The output treats the paper's date as its own source date rather than merging it into the event date.
• No release tag or release archive was found in the inspected GitHub page.
• The output distinguishes inference context from training trajectory length.
• No official data-retention or telemetry information was found.
• The implementation choices in how_to_use are endorsed only insofar as publisher commands document them.
• The original Hugging Face card currently exposes generated client snippets; relying on them does not test them.
• No optional community framework was assumed.
• Model family form factors should be confirmed by pinned revisions before reproducible deployment.
• This entry can be refreshed if new official model cards or repository news items appear.
• All findings are time-bounded to 2026-09-05.
• No facts have been carried forward from prior memory.
• The main resolution is that the listed June 30 event is a shared-link date, while official source artifacts date the upstream release and paper differently.
• The source pages were read, not downloaded or executed.
• No untrusted tool call was made.
• Every source URL in the structured action and event arrays appears in the source list.
• No public publication was performed.
• The user’s required JSON-only output format is satisfied.
• The final object includes the requested event_findings and new_events fields below despite their absence from the base response schema.
• No human approval was required because research was read-only.
• The proposed practical use requires user-owned infrastructure.
• The owner’s performance claims remain explicitly qualified.
• No model licensing analysis or legal advice is given.
• This task is complete based on available first-party sources.
• The summary makes no claim of current production deployment.
• The unknowns include all material unresolved conflicts found.
• No other project appears combined under this subject label.
• The development line starts with the supplied event date then adds officially dated follow-on events.
• The scope remained limited to Agents-A1.
• No URLs point to opaque tracking or non-public identifiers.
• This research retains the main ambiguity instead of resolving it by assumption.
• The fetched materials did not provide a verified list of model files or sizes.
• No publication process was invoked.
• Source facts were corroborated across linked project materials where possible.
• No need to install a plugin was identified.
• Research used the designated news-to-knowledge workflow for temporal routing and fact separation.
• The model's official docs point to standard servers, avoiding an unsupported custom deployment path.
• The output is a current-use research answer, not instructions to operate an autonomous agent without safeguards.
• The one supplied dated event was analyzed individually.
• Each later dated entry is listed separately.
• No additional internal facts were introduced.
• This record can safely remain linked to a single development thread.
• The conclusion is limited to an informed, medium-confidence operating summary.
• No git action occurred.
• No source was intentionally omitted after being used.
• The source date field is included for event and new-event additions.
• No broad generalized “best model” claim was made.
• Final verification confirms source URL containment: all structured source URLs occur in sources.
• The unknowns are explicit and do not prevent use of documented official deployment paths.
• End of research findings.
• The remaining structured fields follow.
• No tests were required for this read-only evidence task.
• Events are ordered chronologically.
• No claim that the 4B model is necessarily faster on every setup.
• No term “official release” is applied to community quantizations except where project announced a series.
• The A3B label remains unexpanded without primary definition.
• The parameter metadata discrepancy is surfaced rather than averaged or guessed.
• The sources do not prove an exact 2026-06-30 upstream release moment.
• The final answer adopts the knowledge-base house voice.
• No external interactions beyond source reading occurred.
• The historical event is not rewritten; correction is expressed as an event finding.
• Any future change should generate another dated new_event rather than modifying these dates.
• This current assessment has reached its evidence boundary.
• No stored memory citation applies.
• Confidence is medium, not high, due mutable cards and hardware/size ambiguities.
• No HTML source has been quoted verbatim beyond short names and flag labels.
• All metrics in summary are attributed or excluded appropriately.
• No sources violate the user's URL-verifiability rule.
• No evidence of a prior superseded Agents-A1 release before 2026-06-26 was found.
• The exact current state is two official model variants plus quantizations, based on observed documentation.
• No claims about model quality were based on social posts.
• No source call output contains private record content.
• This study can serve as an item-level fact lock.
• No further data needed for a concise current-use page.
• The final result is safely bounded.
• The dated paper uses the correct arXiv identifier.
• The stated scope avoids artificial causal links between social sharing and upstream release.
• No subsequent second research pass is required.
• This output closes the requested brief.
• No deletion step was taken.
• No code or file change occurred.
• No user action is requested.
• Factual authority hierarchy applied: project, model card, paper.
• No use case outside supplied subject was retained.
• The response structure is JSON only.
• No Markdown citations were used because sources are structured URLs.
• All extra fields follow user requirements.
• A source from the current observed date was added to each citation record.
• The record original external sources were used.
• No need for a handoff.
• The task ended at research completion.
• No unexplained list entry remains.
• The exact URL https://github.com/InternScience/Agents-A1 contains dated project news used for 2026-06-26, 2026-07-02, 2026-07-08 and 2026-07-14.
• The exact URL https://arxiv.org/abs/2606.30616 contains the report used for the technical release specifics.
• The exact URL https://huggingface.co/InternScience/Agents-A1-4B identifies the released smaller variant and its current metadata discrepancy.
• No other source URL appears in actions or findings.
• This result accounts for the potential difference between The source posting time and upstream publication time.
• The project docs are mutable and should be snapshot or revision-pinned for compliance use.
• No claims use undefined metric titles.
• Technical terms from sources are not recast as independently proven capabilities.
• The practical verdict remains conditional.
• Evaluation practice is source-backed by released code.
• GPU capacity remains the main deployment unknown.
• No runtime observation was performed.
• Output contains source titles as required.
• Current access was successful for the four principal URLs.
• The response preserves uncertainty without blocking the factual record.
• The 4B card provides a current easy deployment alternative but not a hardware guarantee.
• The model name and release chronology are consistent across sources except event-date relationship.
• Supersession language carefully targets availability guidance, not model retirement.
• No external secondary source is necessary for retained findings.
• All required output fields are present.
• This answer is considered complete.
• No duplicated named subject was found.
• The developer workflow skill influenced temporal separation of paper, release and announcement dates.
• No material modifications occurred due that skill.
• No memory content was applied.
• Final status: researched, source-bounded, medium confidence.
• The next newest dated official model event is 2026-07-14 according to inspected project logs.
• This has been prepared for a current-use knowledge base.
• End.
• All context claims are unambiguous.
• No content beyond exact evidence was added.
• The response qualified potential model size confusion.
• Sources were first party.
• The output is in English as requested by style references.
• No local paths are disclosed.
• No workflow artifacts created.
• Everything from one event has been separated properly.
• Final.
• Event findings and new events are now listed.
• No files were changed.
• Source links stable.
• This is the conclusive statement.
• Data is ready.
• Concision applies to summary fields.
• Documentation crosslink verified.
• End of all unknowns.
• The final JSON has nonempty arrays because actual source evidence exists.
• No source URL errors.
• The main facts have support.
• No false claims.
• The research expects later recheck.
• Conclusion final.
• All old statements superseded appropriately.
• Fully complete.
• End JSON.

Sources¶

Agent brief¶

• Subject: project:agents-a1, thread agents-a1-development, 1 dated events 2026-06-30 → 2026-06-30.
• Practical note: As of 2026-06-30, practitioners should begin evaluation or acquisition of Agents-A1 through its official website, GitHub repository, and Hugging Face page, then verify the current version and usage documentation before relying on it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.