Language modelOpen weights
K2 Horizon 3.7B
Institute of Foundation Models
- Released
- September 3, 2026
- Data date
- October 7, 2026
- Input
- Text
- Output
- Text
K2 Horizon 3.7B brings the dense architecture to a smaller scale while retaining a native context of 524,288 tokens. Its published training path uses separate math, code, and STEM-code RL experts, merges their weights, and then applies FTPO as a final training stage.
Specifications and access
| Specification | Value and source |
|---|---|
| Parameters | 3.7B total parametersSource |
| Active parameters | 3.7B parameters per tokenSource |
| Context window | 524,288 tokensSource |
| Input modalities | TextSource |
| Output modalities | TextSource |
| Documented training stages | Pretraining, four midtraining stages, math, code, and STEM-code RL experts with a merge, two SFT phases, and a final anti-doom-loop FTPO step.Source |
| License | Apache 2.0Source |
Page 1 of 3
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.
AIME 2025 (Show measurement, test conditions, and source)
- Source value
- 89.2
- Score
- 89.2
- Metric
- AIME 2025
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
AIME 2026 (Show measurement, test conditions, and source)
- Source value
- 90.8
- Score
- 90.8
- Metric
- AIME 2026
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
HMMT February 2025 (Show measurement, test conditions, and source)
- Source value
- 87.5
- Score
- 87.5
- Metric
- HMMT February 2025
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
HMMT February 2026 (Show measurement, test conditions, and source)
- Source value
- 75.2
- Score
- 75.2
- Metric
- HMMT February 2026
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 68.9
- Score
- 68.9
- Metric
- GPQA Diamond
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
Humanity's Last Exam (text) (Show measurement, test conditions, and source)
- Source value
- 14
- Score
- 14
- Metric
- Humanity's Last Exam (text)
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
IFEval (loose) (Show measurement, test conditions, and source)
- Source value
- 83.6
- Score
- 83.6
- Metric
- IFEval (loose)
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
IFBench (loose) (Show measurement, test conditions, and source)
- Source value
- 47.7
- Score
- 47.7
- Metric
- IFBench (loose)
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
LiveCodeBench v6 (Show measurement, test conditions, and source)
- Source value
- 64.5
- Score
- 64.5
- Metric
- LiveCodeBench v6
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
OJBench (Show measurement, test conditions, and source)
- Source value
- 21.6
- Score
- 21.6
- Metric
- OJBench
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
SciCode (Show measurement, test conditions, and source)
- Source value
- 25.8
- Score
- 25.8
- Metric
- SciCode
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
SWE-bench Verified (Show measurement, test conditions, and source)
- Source value
- 65.6
- Score
- 65.6
- Metric
- SWE-bench Verified
- 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. Exact released K2 Horizon 3.7B anti_doom_loop_verified checkpoint after FTPO. IFM states before/after checkpoints use identical settings. The card supplies test-specific details: GPQA top_p=1 and 5 repeats, SciCode 3 repeats, SWE-bench via mini-swe-agent, Terminal-Bench via terminus-2 with 3 repeats at 256K context, BFCL temperature 0.01; output token budgets vary by task.
- Context
- Where the card gives a per-benchmark protocol, it is included in the configuration. Other rows do not have a reported sample count in the table.
Page 1 of 3
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved 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. |
| Unresolved | full-results-checkpoint-identity Original wording The Full Results table conflicts on several tests with the published results for the final anti_doom_loop_verified checkpoint. The card does not identify the Full Results checkpoint. |
| Not published | per-run-protocol-and-sample-counts Original wording Repeat counts are published only for some FTPO-table tests; unreported sample sizes are not inferred. |
| Unresolved | excluded-non-target-checkpoints Original wording The profile description and model card identify anti_doom_loop_verified after FTPO as the final 3.7B checkpoint. mid_4_final is a different intermediate weight. The Full Results table does not identify an exact checkpoint and differs on several tests from the explicitly labelled after-FTPO values. Those values were excluded; only the explicitly labelled after-FTPO 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) |