Search COMPUTER USE

Search the evidence-backed index by name, vendor, or use case.

COMPUTER USE / PROFILE

Microsoft Fara1.5: capabilities, evidence & fit

An open-weight browser-action model family with screenshot-only perception and a caller-owned execution loop.

Computer-use modelopen source

Fara1.5 supplies the action model, with separate 4B, 9B and 27B checkpoints. Use the documented Fara/MagenticLite harness to execute browser actions; this is not a managed browser subscription. The 9B card is the detailed control-surface reference here.

Official website
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

academic

Fara1.5 comprises 4B, 9B and 27B browser-agent models based on Qwen3.5.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

academic

The agent uses recent screenshots and emits actions sequentially; permitted actions include browser control, search and user questions.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

academic

The research article was released May 21, 2026; its July 22 update reports MIT-licensed weight availability.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

official repo

Microsoft publishes separate Fara1.5-4B, Fara1.5-9B and Fara1.5-27B cards, each declaring MIT. These are exact 1.5 identities, not the older Fara-7B model.

Verified Sep 28, 2026 · verified

huggingface.co

verified · License declaration verified on model cards; no legal interpretation or third-party dependency audit.

official repo

The 9B Files view lists four safetensors shards plus an index and configuration files. No model weights were downloaded or executed.

Verified Sep 28, 2026 · verified

huggingface.co

verified · File existence does not verify checksum, operability or weights for the 4B/27B variants.

official repo

The 9B card documents screenshot-only observations rather than DOM/accessibility input, a 262,144-token context, Qwen3.5-9B base, and a caller-executed browser loop with MagenticLite/Fara CLI guidance.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

official repo

The card describes trained pause behavior and recommends sandbox/allowlist/oversight controls. Training claims do not prove enforced approval or security guarantees in every integration.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

official repo

Exact 4B/9B/27B MIT-declared cards found; 9B four weight shard filenames visible. No weight download, execution or standalone-license/dependency audit. Use exact 1.5 card identity; keep 4B/27B weight inventories and operational validation unverified.

Verified Sep 28, 2026 · partially verified

huggingface.co

conflicted · Conflicting or incompletely resolved evidence is not used as a numerical current value.

official repo

Microsoft research article in the original ledger says 88.6; exact model card says 89.3. No matching run/version explanation found. Preserve original result observation and new conflicting card; no score substitution or formal result.

Verified Sep 28, 2026 · partially verified

huggingface.co

conflicted · Conflicting or incompletely resolved evidence is not used as a numerical current value.

academic

Official current documentation and product sources identify Microsoft Fara1.5 in the scope described here; account access and operational readiness have not been independently tested.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

Last verified
Sep 28, 2026
Primary sources
4
Evidence state
Fresh
Benchmark records
No linked results

Where it fits

CONSIDER IT FOR
  • Teams evaluating self-operated browser model weights
  • Experiments that can pin a checkpoint and Fara-compatible harness
LOOK ELSEWHERE WHEN
  • You need a turnkey hosted automation service
  • You require a proven security guarantee from trained pause behavior

Fit recommendations are editorial interpretations of the documented architecture.

Documented capabilities

CAPABILITYDOCUMENTED STATESOURCE
model familyFara1.5 comprises 4B, 9B and 27B browser-agent models based on Qwen3.5.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

academic

Fara1.5 comprises 4B, 9B and 27B browser-agent models based on Qwen3.5.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

interactionThe agent uses recent screenshots and emits actions sequentially; permitted actions include browser control, search and user questions.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

academic

The agent uses recent screenshots and emits actions sequentially; permitted actions include browser control, search and user questions.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

release historyThe research article was released May 21, 2026; its July 22 update reports MIT-licensed weight availability.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

academic

The research article was released May 21, 2026; its July 22 update reports MIT-licensed weight availability.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

exact model identityMicrosoft publishes separate Fara1.5-4B, Fara1.5-9B and Fara1.5-27B cards, each declaring MIT. These are exact 1.5 identities, not the older Fara-7B model.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

official repo

Microsoft publishes separate Fara1.5-4B, Fara1.5-9B and Fara1.5-27B cards, each declaring MIT. These are exact 1.5 identities, not the older Fara-7B model.

Verified Sep 28, 2026 · verified

huggingface.co

verified · License declaration verified on model cards; no legal interpretation or third-party dependency audit.

weight inventory 9bThe 9B Files view lists four safetensors shards plus an index and configuration files. No model weights were downloaded or executed.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

official repo

The 9B Files view lists four safetensors shards plus an index and configuration files. No model weights were downloaded or executed.

Verified Sep 28, 2026 · verified

huggingface.co

verified · File existence does not verify checksum, operability or weights for the 4B/27B variants.

vision runtime 9bThe 9B card documents screenshot-only observations rather than DOM/accessibility input, a 262,144-token context, Qwen3.5-9B base, and a caller-executed browser loop with MagenticLite/Fara CLI guidance.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

official repo

The 9B card documents screenshot-only observations rather than DOM/accessibility input, a 262,144-token context, Qwen3.5-9B base, and a caller-executed browser loop with MagenticLite/Fara CLI guidance.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

safety boundaryThe card describes trained pause behavior and recommends sandbox/allowlist/oversight controls. Training claims do not prove enforced approval or security guarantees in every integration.
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

official repo

The card describes trained pause behavior and recommends sandbox/allowlist/oversight controls. Training claims do not prove enforced approval or security guarantees in every integration.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

Environment, deployment & control

Environment

browser
Full
desktop
Unknown
mobile
Unknown

Deployment

cloud
Unknown
local
Unknown
self Hosted
Unknown

Extensibility

api
Unknown
sdk
Unknown
mcp
Unknown
bring Your Own Model
Unknown

Control method

screenshots
Full
mouse Keyboard
Full
dom Or Accessibility
Unknown
code Execution
Unknown

Unknown = insufficient public evidence; No evidence = an explicit negative record; N/A = does not apply to this layer.

Trust, oversight & limitations

Documented oversight

human Approval
Unknown
audit Logs
Unknown
sandboxing
Unknown
prompt Injection Defense
Unknown
credential Controls
Unknown
EVIDENCE RECORD

Follow the source.

Source verification confirms what the publisher states. It is not our independent performance test.

academic

Fara1.5 comprises 4B, 9B and 27B browser-agent models based on Qwen3.5.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

academic

The agent uses recent screenshots and emits actions sequentially; permitted actions include browser control, search and user questions.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

academic

The research article was released May 21, 2026; its July 22 update reports MIT-licensed weight availability.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

official repo

Microsoft publishes separate Fara1.5-4B, Fara1.5-9B and Fara1.5-27B cards, each declaring MIT. These are exact 1.5 identities, not the older Fara-7B model.

Verified Sep 28, 2026 · verified

huggingface.co

verified · License declaration verified on model cards; no legal interpretation or third-party dependency audit.

official repo

The 9B Files view lists four safetensors shards plus an index and configuration files. No model weights were downloaded or executed.

Verified Sep 28, 2026 · verified

huggingface.co

verified · File existence does not verify checksum, operability or weights for the 4B/27B variants.

official repo

The 9B card documents screenshot-only observations rather than DOM/accessibility input, a 262,144-token context, Qwen3.5-9B base, and a caller-executed browser loop with MagenticLite/Fara CLI guidance.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

official repo

The card describes trained pause behavior and recommends sandbox/allowlist/oversight controls. Training claims do not prove enforced approval or security guarantees in every integration.

Verified Sep 28, 2026 · verified

huggingface.co

verified · Provider documentation or selected billing-mode observation; no independent product execution.

official repo

Exact 4B/9B/27B MIT-declared cards found; 9B four weight shard filenames visible. No weight download, execution or standalone-license/dependency audit. Use exact 1.5 card identity; keep 4B/27B weight inventories and operational validation unverified.

Verified Sep 28, 2026 · partially verified

huggingface.co

conflicted · Conflicting or incompletely resolved evidence is not used as a numerical current value.

official repo

Microsoft research article in the original ledger says 88.6; exact model card says 89.3. No matching run/version explanation found. Preserve original result observation and new conflicting card; no score substitution or formal result.

Verified Sep 28, 2026 · partially verified

huggingface.co

conflicted · Conflicting or incompletely resolved evidence is not used as a numerical current value.

academic

Official current documentation and product sources identify Microsoft Fara1.5 in the scope described here; account access and operational readiness have not been independently tested.

Verified Sep 28, 2026 · verified

www.microsoft.com

verified · Documented by the provider; not an independent execution, reliability or security test.

Know the boundary

  • The 27B WebVoyager figure conflicts across Microsoft sources; no score is imported.
  • Weights were not downloaded or executed by COMPUTER USE. 4B/27B file inventory and standalone license/dependency review remain incomplete.
  • Training for approval behavior is not enforcement; isolate the runtime and enforce consequential-action approval in code.
Design human approval →

Pricing snapshot

Price needs verification

Current pricing is not sufficiently verified, or its 14-day review window has expired. Check the official provider before budgeting.

Evidence, scope & unresolved questions

Stack classification: L1. Verified means the cited source supports the statement; it does not mean COMPUTER USE independently tested the product.

Change & verification history

RESEARCH NOTES

Sources & verification

Reviewed Sep 28, 2026. Architecture recommendations are editorial analysis; linked vendor documentation supports the underlying capability and safety facts.

  1. www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent
  2. huggingface.co/microsoft/Fara1.5-4B
  3. huggingface.co/microsoft/Fara1.5-9B/tree/main
  4. huggingface.co/microsoft/Fara1.5-9B
How we verify evidence →