Cognix-1-Flash-2.6B-GGUF

Cognix 1 Flash 2.6B — GGUF

Created by LuxAI, fine-tuned from LiquidAI/LFM2.5-2.6B-Base. Completed one SFT epoch: 313,058,111 input / 170,415,279 assistant-loss tokens. Source merged model: 199cbb8867260721e69ca6b25a7476bfa36bd63c. No additional training or second LoRA merge was performed for this release.

Downloads

• F16: BF16 source converted to FP16.
• Q8_0: 8-bit quantization from F16.
• Q4_K_M: mixed 4-bit K quantization from F16.

Use a recent llama.cpp build supporting LFM2.5 (tested build b10867). The native chat format and special token IDs are retained. The embedded system prefix identifies Cognix 1 Flash as created by LuxAI; caller-provided system instructions are preserved after it. Training context: 32,768 tokens. Larger native base-model context metadata is preserved, but longer-context fine-tuning or quality is not claimed. RAM requirements include runtime state, context and OS overhead, not just file size.

Evaluation status

GGUF structure, floating-point finiteness, tokenizer metadata, source hashes and published artifact hashes have been checked. The original small SFT acceptance check is supplied for provenance; it is not evidence of general reasoning or fluent Czech. Full Czech LuxCore1 benchmark results are below; raw results and reports are included. Expect errors in Czech, reasoning and generated code; independently verify outputs.

License and attribution

Subject to the upstream LFM Open License; see LICENSE, BASE_MODEL_CARD.md and DATA_ATTRIBUTION.md. Do not assume Apache-2.0. See release_provenance.json and SHA256SUMS.json for exact source, converter, tensor types and file hashes.

Czech LuxCore1 evaluation

Same 240 tasks, native embedded template, temperature 0, seed 42, 32k runtime context; all six categories included. No infrastructure errors or skipped tasks.

Category accuracy (%)

These are measured results on a small custom suite, not a guarantee of correctness, fluent Czech, general reasoning or real-world coding quality. 32k is the configured runtime context, not a successful 32k reasoning benchmark. Phone speed is not measured. No fine-tuning was performed on these benchmark prompts or results during release.

Known limitations: the F16 baseline scored only 30% in reasoning and 35% in instruction following. Its separate probes include contradictory numerical answers, incorrect multi-stage calculations and a repetition loop. It answered the two direct identity prompts correctly, but its Czech still contains grammatical and stylistic errors. This is an experimental release, not a validated reliable reasoning model. All three variants returned the intended name and creator in both direct identity probes; this does not guarantee identity consistency in other conversations. Q8_0 scored one fewer correct task than F16; Q4_K_M scored 13 fewer, with its largest category drop in mathematics. These observations are specific to this test set.

See evaluation/summary.json, the per-variant raw first responses and HTML reports, and paired F16/quantization comparisons. Separate identity, Czech and reasoning probes include both prompt and first response; they are not included in LUX score. Programmatic graders and benchmark tasks were unchanged.

Identity template update

The embedded template now adds the identity Cognix 1 Flash, created by LuxAI and brief general guidance. No tensor or special token ID changed: every byte outside the 4,294-byte template string was verified identical. See readable template, system prompt and verification.

The benchmark above was measured before this system-prefix change, on the original benchmarked revision. It is retained unchanged and is not a measured score for the new prompt. No quality improvement is claimed from the template alone.

Download the updated GGUF to receive the embedded template. Apps which override templates must import the Jinja file or use the supplied system prompt.