Skip to content

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

NameLast commit date
..
_eq.py
_gen.py
_plots.py
activation_spikes.png
attn-res.mmd
attn-res.png
belady_vs_lru.png
big-picture.mmd
big-picture.png
binary_sizes.png
build-flow.mmd
build-flow.png
bytes_census.png
bytes_paradox.png
cache_wakeup.png
cache-3phase.mmd
cache-3phase.png
cache-slot-states.mmd
cache-slot-states.png
config-guard.mmd
config-guard.png
context_scaling.png
decoder-layer.mmd
decoder-layer.png
duplicate_config.png
eq_absmax.png
eq_accum_order.png
eq_align.png
eq_attn_res.png
eq_bf16_widen.png
eq_block_count.png
eq_bytes_per_weight.png
eq_compulsory.png
eq_context_ceiling.png
eq_dequant_cost.png
eq_fit_ledger.png
eq_gb_per_layer.png
eq_io_share.png
eq_kda_decay.png
eq_kda_gate.png
eq_kda_recurrence.png
eq_kda_state.png
eq_kv_compression.png
eq_l2norm.png
eq_layer_map.png
eq_mla_kv.png
eq_mla_latent.png
eq_moe_latent.png
eq_mxfp4.png
eq_naive_memory.png
eq_nibble_pack.png
eq_prefetch.png
eq_prefill.png
eq_queue_depth.png
eq_ram_budget.png
eq_resident_set.png
eq_retention.png
eq_rmsnorm.png
eq_router.png
eq_shortconv.png
eq_situ_glu.png
eq_slot_count.png
eq_sparsity_ratio.png
eq_spread.png
eq_st_layout.png
eq_stream_vs_quantize.png
eq_tolerance.png
eq_traffic.png
eq_trunk_hit.png
eq_working_set.png
expert_reuse.png
first-token.mmd
first-token.png
fit_cascade.png
forward-pass.mmd
forward-pass.png
gate_ladder.png
gb_read_shelf.png
gen-loop.mmd
gen-loop.png
hit_metrics.png
hosts-confound.mmd
hosts-confound.png
io_share.png
kda-flow.mmd
kda-flow.png
kda-nine-steps.mmd
kda-nine-steps.png
kda-state.mmd
kda-state.png
kernel-contract.mmd
kernel-contract.png
kv_layout.png
ladder-harness.mmd
ladder-harness.png
layer_conformance.png
machine-model.mmd
machine-model.png
main_architecture_with_spongbob.png
main_architecture.png
memory_ladder.png
metric-bug.mmd
metric-bug.png
mla-latent.mmd
mla-latent.png
moe-dispatch.mmd
moe-dispatch.png
moe-sparsity.mmd
moe-sparsity.png
mxfp4_savings.png
mxfp4-decode.mmd
mxfp4-decode.png
noise-floor.mmd
noise-floor.png
oracle-ladder.mmd
oracle-ladder.png
pack_progress.png
pack-trunk.mmd
pack-trunk.png
parity-flow.mmd
parity-flow.png
patrick_pray.png
prefill_cost.png
preset_ladder.png
quant_by_layer_type.png
quant_error.png
recap.mmd
recap.png
replication_noise.png
resident_vs_streamed.png
roundtrip_sizes.png
rss_fidelity.png
same_answer.png
shard_sizes.png
sim_vs_measured.png
split-rule.mmd
split-rule.png
st-load.mmd
st-load.png
step_trace.png
storage_regimes.png
tok-flow.mmd
tok-flow.png
token_time_split.png
trace-replay.mmd
trace-replay.png
trunk_cache_split.png
trunk-stream.mmd
trunk-stream.png
whats_missing.png