nafanya-0.8b API & Inference Endpoint

What it is

Base: Qwen/Qwen3.5-0.8B (Apache-2.0).
Training: MLX-LM BF16 LoRA, 2 last layers, rank 16, scale 16, dropout 0.05, 75 iterations (the run was stopped by the macOS GPU watchdog), batch 1, max sequence 1280.
Data: 831 synthetic Russian requests and refusals (v5 + targeted v7), generated by DeepSeek V4.1 Flash (AI-generated), filtered against the agent's frozen test sets.
Same compact prompt and 13 tools as nafanya-1.7b.

Evaluation

temperature 0, compact prompt, GGUF. The model alone is weak on fresh phrasings; the agent's deterministic rules layer is what makes the combination usable.

Table with columns: Set, Model alone, With kuzya rules
Set	Model alone	With kuzya rules
held-out v4 (fresh)	25/40 (62%), args 22/40	39/40 (98%), args 39/40
held-out v3	—	38/40 (95%)
held-out v2	—	40/40 (100%)
first 30 frozen queries, phone	—	26/30 (87%)

Phone (Poco F1): peak RAM 0.92 GB, load 2 s, warm energy 4.1 J per query with rules, 0.6 s per query.

Do not attribute the rules-assisted numbers to the weights: the pre-routing, refusals and argument repair live in the kuzya agent.

Limitations

Weaker than nafanya-1.7b on free-form requests that the rules do not cover.
On its own it often fails negatives and out-of-scope requests.
Same semantic weak spots as the 1.7B: relative volume/brightness, mute/nosleep inversion, service-log arguments.

Data and licence

Weights: Apache-2.0.
Training data is synthetic, generated by DeepSeek V4.1 Flash via its API under the DeepSeek Open Platform Terms of Service (release 2026-04-22, §4.2(3)). It is AI-generated; no phone data, no personal data.
DeepSeek is mentioned only as the data source; no affiliation or endorsement.

What it is

Base: Qwen/Qwen3.5-0.8B (Apache-2.0).
Training: MLX-LM BF16 LoRA, 2 last layers, rank 16, scale 16, dropout 0.05, 75 iterations (the run was stopped by the macOS GPU watchdog), batch 1, max sequence 1280.
Data: 831 synthetic Russian requests and refusals (v5 + targeted v7), generated by DeepSeek V4.1 Flash (AI-generated), filtered against the agent's frozen test sets.
Same compact prompt and 13 tools as nafanya-1.7b.

Evaluation

temperature 0, compact prompt, GGUF. The model alone is weak on fresh phrasings; the agent's deterministic rules layer is what makes the combination usable.

Table with columns: Set, Model alone, With kuzya rules
Set	Model alone	With kuzya rules
held-out v4 (fresh)	25/40 (62%), args 22/40	39/40 (98%), args 39/40
held-out v3	—	38/40 (95%)
held-out v2	—	40/40 (100%)
first 30 frozen queries, phone	—	26/30 (87%)

Phone (Poco F1): peak RAM 0.92 GB, load 2 s, warm energy 4.1 J per query with rules, 0.6 s per query.

Do not attribute the rules-assisted numbers to the weights: the pre-routing, refusals and argument repair live in the kuzya agent.

Limitations

Weaker than nafanya-1.7b on free-form requests that the rules do not cover.
On its own it often fails negatives and out-of-scope requests.
Same semantic weak spots as the 1.7B: relative volume/brightness, mute/nosleep inversion, service-log arguments.

Data and licence

Weights: Apache-2.0.
Training data is synthetic, generated by DeepSeek V4.1 Flash via its API under the DeepSeek Open Platform Terms of Service (release 2026-04-22, §4.2(3)). It is AI-generated; no phone data, no personal data.
DeepSeek is mentioned only as the data source; no affiliation or endorsement.

nafanya-0.8b

README

What it is

Evaluation

Limitations

Data and licence

Links

Explore FriendliAI today

README

What it is

Evaluation

Limitations

Data and licence

Links