Fijik 2.0 350M SFT

Fijik 2.0 350M SFT is the first, agentic, edge-LLM we have ever released. it is designed specifically for edge devices and quick inference it was trained efficiently on just one GPU. In addition the model can work with up to 65K context (thanks to YARN scaling) and is based off of granite-4.0-350M-base. (non H, aka no mamba layers).

Fijik 2.0 supports the following reasoning efforts, configurable via the chat template: disabled <--- Configures the model to not think at all, it will not generate reasoning tags, best for chat, title generation.

low <--- Configures the model to think a little before responding, good for web search, tool calls etc.

medium <--- Configures the model to think well before responding, good for code etc.

high <--- Configures the model to think hard before responding, may loop, could improve math, code performance, not needed for most users.

We recommend the following sampling parameters:

tempature: 0.7
Top K: 40
Repeat Penalty: 1.05
Presence Penalty: 0.25
Top P Sampling: 0.95
Min P sampling: 0 or disabled

Benchmarks

NOTE: all are pass@4, aka "could the model get it right at least once"

NOTE2: all benchmarks were run internally and using the exact same sampling parameters

Benchmark	Fijik2.0-350M-d	LFM2.5-350M	granite4.0-h-350M	SmolLM2-360M
MATH500	9.20	21.80	29.60	1.40**
GPQA-DIAMOND	65.15	33.33	53.03	errors**
MMLU REDUX*	65.60	47.20	61.07	nill
HUMANEVAL	22.56	17.68	38.41	9.76

For fijik2.0: d efforts: d = disabled to make benchmark(s) fair as other than fijik all the others are non-reasoning models

*mmlu_redux only 10 subjects, 750 items.

**SmolLM2 often fell into loops, not enough context, etc. MMLU not included, math500 had 7% errors and thus is not trustworthy. GPQA had so many errors that it has been removed for smollm2.

In this benchmark table we can see Fijik2.0 performing competitively with similarly sized models, though fijik2.0 is weak on math, probably due to no RL.

Training metholodgy

    Granite 4.0 350M base
      \           /
    Continual pre-training1 (~6B tokens, august 2025 knowledge cutoff, low context)
       \         / 
    Continual pre-training2 (~1B tokens, aug 2025 knowledge cutoff, higher context)
        \       /
    Supervised fine-tuning  (~3B tokens, data from feb 2026, mixed chat, agentic, web search, code, tool use, reasoning)
          \   /
    Final uploaded model    (best checkpoint chosen)

For CPT a 2bit ademamix-style optimizer was used. For SFT a 4bit ademamix-style optimizer was used. Both optimizers worked well. Fijik 2.0 got 1.1 loss at its selected sft checkpoint. (1.3 EPOCH)

Safety

Fijik-2.0 has not undergone RLHF. Although it still is safety trained thanks to SFT. In addition fijik has been trained to state when it does not know something. Though ideally in the system prompt tell it specifically that if it has no information about something, then it should state so.