Language modelOpen weights

K2 Horizon 375B-A23B

Institute of Foundation Models

Released: September 3, 2026
Data date: October 7, 2026

Input: Text
Output: Text

K2 Horizon 375B-A23B is the largest model in IFM’s open fleet. Its MoE architecture combines 375 billion total parameters and activates 23 billion per token. IFM also releases intermediate checkpoints from pretraining, midtraining, reinforcement learning, and SFT, making several stages of its development available for study.

Specifications and access

Specification	Value and source
Parameters	375B total parametersSource
Active parameters	23B parameters per tokenSource
Context window	524,288 tokensSource
Input modalities	TextSource
Output modalities	TextSource
Documented training stages	Two pretraining phases, four midtraining stages, five RL experts followed by a merge, and three SFT phases.Source
License	Apache 2.0Source

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

τ³-bench Banking (Show measurement, test conditions, and source)

Source value: 34
Score: 34
Metric: success rate
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-card result for the 375B-A23B checkpoint; task label is τ³-bench Banking.

Source IFM model card

Toolathlon Verified (Show measurement, test conditions, and source)

Source value: 65.3
Score: 65.3
Metric: score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported result for K2 Horizon 375B-A23B; no per-task sampling or run count is given in the row.

Source IFM model card

AutomationBench Public (Show measurement, test conditions, and source)

Source value: 25.3
Score: 25.3
Metric: success rate
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result on the public AutomationBench split.

Source IFM model card

Apex-Agents pass@1 (Show measurement, test conditions, and source)

Source value: 24.8
Score: 24.8
Metric: pass@1
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result; score is explicitly pass@1 and the card notes the text-only subset.

Source IFM model card

MCPMark (Show measurement, test conditions, and source)

Source value: 67.7
Score: 67.7
Metric: score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result on MCPMark.

Source IFM model card

BrowseComp (Show measurement, test conditions, and source)

Source value: 72.8
Score: 72.8
Metric: score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result; IFM's evaluation note uses Discard-all@95K. The source cautions that BrowseComp harnesses can vary by model.

Source IFM model card

WildClawBench (Show measurement, test conditions, and source)

Source value: 50.9
Score: 50.9
Metric: score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result; source notes the English text-only task subset.

Source IFM model card

Terminal-Bench 2.1, as initially reported (Show measurement, test conditions, and source)

Source value: 70.2
Score: 70.2
Metric: pass rate
Unit: %
Category: source-specific
Direction: Higher is better
Sample: 712
Source type: vendor-reported
Evaluator: IFM K2 Horizon announcement
Status: disputed
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported K2 Horizon 375B-A23B run on Terminal-Bench 2.1, with 89 tasks and 8 attempts per task (712 attempts). The later IFM audit flagged reward-hacking trials.
Context: Do not treat as the audited score: IFM's later analysis flagged 24 trials across 10 tasks for reward hacking and reports 66.9% after removing them.

Source IFM K2 Horizon announcement

Terminal-Bench 2.1, after reward-hacking audit (Show measurement, test conditions, and source)

Source value: 66.9
Score: 66.9
Metric: pass rate
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM K2 Horizon announcement
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM's audit of the 375B-A23B Terminal-Bench 2.1 result using Artificial Analysis Harbor analysis and a reward-hacking criterion; 24 flagged trials removed, 79 tasks retained.
Context: Audited result reported by IFM; the source does not publish the exact post-filter attempt denominator.

Source IFM K2 Horizon announcement

SciCode (Show measurement, test conditions, and source)

Source value: 42.7
Score: 42.7
Metric: score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result on SciCode.

Source IFM model card

SWE-Atlas-QnA, strict (Show measurement, test conditions, and source)

Source value: 48.4
Score: 48.4
Metric: strict score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result using the strict SWE-Atlas-QnA setting; the card notes no internet access.

Source IFM model card

SWE-bench Pro, strict (Show measurement, test conditions, and source)

Source value: 42.6
Score: 42.6
Metric: strict score
Unit: %
Category: source-specific
Direction: Higher is better
Source type: vendor-reported
Evaluator: IFM model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. IFM-reported exact-model result under strict SWE-bench Pro evaluation; the source notes no internet access.

Source IFM model card

Page 1 of 3

Sources and data date

Type	Evidence and data date
Input	Documented modalities. Retrieved October 7, 2026. Source: K2 Horizon 375B-A23B model card · K2 Horizon 375B-A23B model card (retrieved October 7, 2026) · K2 Horizon 375B-A23B model card · Editorial description reviewed October 7, 2026
Output	Documented modalities. Retrieved October 7, 2026. Source: K2 Horizon 375B-A23B model card · K2 Horizon 375B-A23B model card (retrieved October 7, 2026) · K2 Horizon 375B-A23B model card · Editorial description reviewed October 7, 2026
Research status	Research date October 7, 2026. 11 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	per-run-protocol-and-sample-counts Original wording IFM does not publish repeat counts or full harness details for many table entries; these remain unknown.
Unresolved	excluded-non-target-checkpoints Original wording The pinned IFM family table contains values for intermediate checkpoint mid_4_final. Those weights are different from the final 375B-A23B target profile. The midtraining values were therefore excluded; final model-card results remain.
Additional source	IFM K2 Horizon release announcement (retrieved October 7, 2026) · IFM K2 Horizon announcement · Editorial description reviewed October 7, 2026
Additional source	IFM model card (retrieved October 7, 2026)
Additional source	Artificial Analysis model profile (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)
Additional source	Model metadata (retrieved October 7, 2026)