ColGrep-Agent-2B has 2.517 billion total parameters. The selected Q4_K_M GGUF file is 1.561 GB.
Data snapshot: . 4 published tests in this catalog.
Model file and details
Provider
LightOn
Total parameters
2,516,756,480
Exact file size
1,561,318,592 bytes
Quantization / precision
Q4_K_M
Weight repository
lightonai/colgrep-agent-2B-GGUF
Original model repository
lightonai/colgrep-agent-2B
Weight revision
c2e69958d03af9188af9ebf89582f8887f0bf534
Model revision
f798a61129c6decb5166864e0a3c98373e62c764
Runtime notes
Confirm model architecture and quantization support in your runner.
Recorded release date
2026-10-03
Date evidence
Official source repository created 3 October 2026; official GGUF repository created 9 October. Revisions pinned after LightOn corrected the training method from DAPO to GRPO on 11 October; the selected weights remained byte-identical.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
SWE-bench Lite localization · File F1
76.18 / 100
Publisher-reported file-level F1 for code localization in the ColGrep harness: 300 SWE-bench Lite instances per seed, averaged over five seeds, strict predictions without rescue and a 10-turn limit.
Benchmark efficiency: 48.792 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: LightOn publisher evaluation of the source checkpoint in the ColGrep harness on 300 SWE-bench Lite instances per seed, averaged over five seeds. File F1 measures overlap with reference files. The agent had a 10-turn limit with mandatory final submission; predictions were scored strictly without rescue.
Source-checkpoint result, not an independent test and not a benchmark of the selected Q4_K_M file. LightOn corrected the training description from DAPO to GRPO on 11 October 2026; the reported scores and selected GGUF bytes did not change. The model is a specialist that returns relevant file and line locations; it does not generate end-to-end patches. The official ColGrep installer defaults to the larger Q8_0 artifact, not this Q4_K_M. The 1.561 GB file excludes runtime, the ColGrep index, KV cache and context memory and does not establish device compatibility or speed. Apache-2.0.
Publisher-reported line-level F1 for code localization in the ColGrep harness: 300 SWE-bench Lite instances per seed, averaged over five seeds, strict predictions without rescue and a 10-turn limit.
Benchmark efficiency: 12.758 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: LightOn publisher evaluation of the source checkpoint in the ColGrep harness on 300 SWE-bench Lite instances per seed, averaged over five seeds. Line F1 measures overlap with reference lines. The agent had a 10-turn limit with mandatory final submission; predictions were scored strictly without rescue.
Source-checkpoint result, not an independent test and not a benchmark of the selected Q4_K_M file. The result is code-localization accuracy, not issue-resolution rate, token speed or a general coding score. Quantization, ColGrep version, repository index and runtime settings can change local behavior.
Publisher-reported file-level F1 for multilingual code localization in the ColGrep harness: 300 Multi-SWE-bench Flash instances per seed, averaged over five seeds, strict predictions without rescue and a 10-turn limit.
Benchmark efficiency: 40.075 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: LightOn publisher evaluation of the source checkpoint in the ColGrep harness on 300 Multi-SWE-bench Flash instances per seed, averaged over five seeds. File F1 measures overlap with reference files. The agent had a 10-turn limit with mandatory final submission; predictions were scored strictly without rescue.
Source-checkpoint result, not an independent test and not a benchmark of the selected Q4_K_M file. This is a multilingual code-localization result, not issue-resolution rate, token speed or a general coding score. Quantization, ColGrep version, repository index and runtime settings can change local behavior.
Publisher-reported line-level F1 for multilingual code localization in the ColGrep harness: 300 Multi-SWE-bench Flash instances per seed, averaged over five seeds, strict predictions without rescue and a 10-turn limit.
Benchmark efficiency: 10.203 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: LightOn publisher evaluation of the source checkpoint in the ColGrep harness on 300 Multi-SWE-bench Flash instances per seed, averaged over five seeds. Line F1 measures overlap with reference lines. The agent had a 10-turn limit with mandatory final submission; predictions were scored strictly without rescue.
Source-checkpoint result, not an independent test and not a benchmark of the selected Q4_K_M file. This is a multilingual code-localization result, not issue-resolution rate, token speed or a general coding score. Quantization, ColGrep version, repository index and runtime settings can change local behavior.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 1.561 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 1.56 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-03. Apple US launch price announced 22 September 2026: USD 1,699. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Confirm model architecture and quantization support in your runner. The selected weights use 1.561 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.