Inference · /inferences/seasons/2026-W43

Season 2026-W43, open.

19 Oct 2026 to 26 Oct 2026 — 6 days left. What the grams proved of the catalog this window, per artifact and hardware class. Lineage is from the catalog; every number is a signed measurement or a count of them, and a gram count is a verified floor.

2grams contributed

27artifacts proved

1hardware classes

10publishers

cohorts this season

0 · Lead member, rung 09
1 · Lead member's ladder17
2 · Other vendor members1
3 · Derivatives0

The order a season proves things in: the number every other number is compared to comes first.

hardware classes that signed

Linux · x86-64 · NVIDIA · 32 GiB

A class is os · arch · card class · RAM band. A number proved on one says nothing about another.

publishers, by what the network proved of their builds

unsloth9 held probes · 2 grams · ollama 9
Qwen6 held probes · 1 gram · ollama 5, vllm 1
bartowski5 held probes · 1 gram · ollama 5
UsefulSensors2 held probes · 1 gram · speech-pod 2
google1 held probe · 1 gram · ollama 1
kokoro1 held probe · 1 gram · speech-pod 1, speech-host 1
MaziyarPanahi1 held probe · 1 gram · ollama 1
mradermacher1 held probe · 1 gram · ollama 1
QuantFactory1 held probe · 1 gram · ollama 1
sarvamai1 held probe · 1 gram · vllm 1

Listed by held outcomes. Downloads, likes and popularity do not appear on this page, because nobody signs them.

cells the network has signed as unprovable

bharatgenai/sooktam2Windows · x86-64 · Intel GPU · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
coqui/XTTS-v2Windows · x86-64 · NVIDIA · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
coqui/XTTS-v2macOS · ARM64 · Apple GPU · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
coqui/XTTS-v2Windows · x86-64 · NVIDIA · 8 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
hf.co/unsloth/gemma-4-E2B-it-GGUF:IQ3_XXSLinux · x86-64 · NVIDIA · 32 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
hf.co/unsloth/gemma-4-E2B-it-qat-mobile-GGUF:Q2_K_XLLinux · x86-64 · NVIDIA · 32 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
ResembleAI/chatterboxmacOS · ARM64 · Apple GPU · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
SWivid/F5-TTSWindows · x86-64 · NVIDIA · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream
SWivid/F5-TTSmacOS · ARM64 · Apple GPU · 16 GiB · unavailable · 1 gram
the ref, or the quant asked of it, is not published upstream

refusals are signed facts about (artifact, class): gated, unloadable, or absent upstream. They are coverage of a kind -- the fleet learned it once -- and they sink the cell to the back of every gram of that class. A refusal lifts when its season closes and the engine that refused has moved on.

what proving costs, per hardware class

Windows · x86-64 · Intel GPU · 16 GiB580.6 MiB/s median fetch
- generation: 6 min 3 s median probe · 1 sample
carried forward — no gram of this class reported in the latest aggregate
Windows · x86-64 · NVIDIA · 16 GiB63.6 MiB/s median fetch
- ollama: 17 min 42 s median probe · 1 sample
- ollama-host: 46 min 44 s median probe · 1 sample
- generation: 0 s median probe · 1 sample
macOS · x86-64 · Apple GPU · 8 GiB2.8 MiB/s median fetch
carried forward — no gram of this class reported in the latest aggregate
Windows · x86-64 · AMD · 8 GiB894.6 MiB/s median fetch
- ollama-host: 30 min 52 s median probe · 3 samples
- generation: 1 s median probe · 1 sample
carried forward — no gram of this class reported in the latest aggregate
macOS · ARM64 · Apple GPU · 16 GiB11.4 MiB/s median fetch
- ollama-host: 4 min 8 s median probe · 1 sample
Windows · x86-64 · NVIDIA · 8 GiB13.8 MiB/s median fetch
Windows · x86-64 · no GPU · 8 GiB7.3 MiB/s median fetch
- generation: 1 s median probe · 1 sample
- ollama-host: 30 min 52 s median probe · 3 samples
carried forward — no gram of this class reported in the latest aggregate
Windows · x86-64 · no GPU · 16 GiB5 MiB/s median fetch
carried forward — no gram of this class reported in the latest aggregate
macOS · ARM64 · Apple GPU · 8 GiB9 MiB/s median fetch
- ollama-host: 30 min 18 s median probe · 3 samples
Linux · x86-64 · NVIDIA · 32 GiB109.7 MiB/s median fetch
- speech-pod: 1 min 4 s median probe · 2 samples

cost is operational: medians of node-reported pull/probe clocks per hardware class and engine, with sample counts. It is not a measurement of any model and is never ranked.

what the network actually used

gemma4:e2b223.4k tokens served
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_M192.4k tokens served
claude-haiku-4-5188.3k tokens served
hf.co/unsloth/Qwen3-4B-GGUF:IQ3_XXS152.6k tokens served
hf.co/unsloth/Qwen3-4B-GGUF:IQ1_S127.6k tokens served
hf.co/QuantFactory/OpenHathi-7B-Hi-v0.1-Base-GGUF:Q4_K_M105.3k tokens served
hf.co/unsloth/Phi-4-mini-reasoning-GGUF:IQ1_S84.3k tokens served
hf.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M65.3k tokens served
llama3.2:1b49.6k tokens served
hf.co/Qwen/Qwen3-4B-GGUF:Q5_041.8k tokens served
hf.co/lmstudio-community/Phi-4-mini-reasoning-GGUF:Q8_037.5k tokens served
hf.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_k_m32.2k tokens served

demand is tokens produced per artifact per class, from signed throughput observations; it orders depth, never existence.

Cohort 0 · Lead member, rung 0 · 9

gemma4:e2bollamaQ4_K_Mrung 0
google · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 105.23 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 31 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.bn, lang.en, lang.gu, lang.hi, lang.kn, lang.ml, lang.mr, lang.pa, lang.ta, lang.te, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated, vision, vision.count, vision.multi_image, vision.read, vision.spatial
served 223.4k tokens across the network
hf.co/bartowski/krutrim-ai-labs_Krutrim-2-instruct-GGUF:Q4_K_MollamaQ4_K_Mrung 0
bartowski · quantization of krutrim-ai-labs/Krutrim-2-instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 13.56 tok/s best · trained 1000K · proved 64K ctx (the model lost the needle at 128K) · 15 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, instruct.follow, lang.bn, lang.en, lang.gu, lang.kn, lang.ml, lang.mr, lang.pa, lang.ta, lang.te, output.long, output.schema, research.synthesize, structured_output
served 1.8k tokens across the network
hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_MollamaQ4_K_Mrung 0
bartowski · quantization of meta-llama/Llama-3.2-3B-Instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 86.87 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 15 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, faithful_tool_result, lang.en, lang.hi, lang.kn, lang.te, output.long, output.schema, research.synthesize, structured_output, tools, tools.call, tools.faithful, tools.loop_ends, tools_loop_terminated
served 2.6k tokens across the network
hf.co/bartowski/sarvam-1-GGUF:Q4_K_MollamaQ4_K_Mrung 0
bartowski · quantization of sarvamai/sarvam-1
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 111.22 tok/s best · trained 8K · proved 4K ctx (the model lost the needle at 8K) · 8 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: lang.bn, lang.hi, lang.kn, lang.ml, lang.mr, output.long, output.schema, structured_output
served 3.9k tokens across the network
hf.co/mradermacher/Airavata-GGUF:Q4_K_MollamaQ4_K_Mrung 0
mradermacher · quantization of ai4bharat/Airavata
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 31.59 tok/s best · 4 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: instruct.follow, lang.en, lang.hi, output.long
served 20.4k tokens across the network
hf.co/QuantFactory/OpenHathi-7B-Hi-v0.1-Base-GGUF:Q4_K_MollamaQ4_K_Mrung 0
QuantFactory · quantization of sarvamai/OpenHathi-7B-Hi-v0.1-Base
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 29.84 tok/s best · on host
served 105.3k tokens across the network
hf.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_k_mollamaQ4_K_Mrung 0
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 239.08 tok/s best · trained 32K · proved 32K ctx (recalled at every size asked) · 11 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, faithful_tool_result, instruct.follow, lang.en, output.schema, structured_output, tools, tools.call, tools.faithful, tools.loop_ends, tools_loop_terminated
served 32.2k tokens across the network
hf.co/Qwen/Qwen3-4B-GGUF:Q4_K_MollamaQ4_K_Mrung 0
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 50.44 tok/s best · trained 40K · proved 16K ctx (the model lost the needle at 32K) · 16 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.en, lang.ta, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.multi_hop
served 22.8k tokens across the network
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_MollamaQ4_K_Mrung 0
unsloth · quantization of google/gemma-4-E2B-it
- Linux · x86-64 · NVIDIA · 32 GiB — 2 grams · 77.58 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 54 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.ar, lang.as, lang.bn, lang.de, lang.en, lang.es, lang.fa, lang.fr, lang.gu, lang.he, lang.hi, lang.id, lang.it, lang.ja, lang.kn, lang.ko, lang.ml, lang.mr, lang.ms, lang.ne, lang.or, lang.pa, lang.pl, lang.pt, lang.si, lang.ta, lang.te, lang.th, lang.tr, lang.uk, lang.ur, lang.vi, lang.zh, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated, vision, vision.count, vision.multi_image, vision.read, vision.spatial
served 192.4k tokens across the network

Cohort 1 · Lead member's ladder · 17

hf.co/Qwen/Qwen3-4B-GGUF:Q8_0ollamaQ8_0rung 1
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 38.86 tok/s best · trained 40K · proved 32K ctx (recalled at every size asked) · 15 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.en, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.multi_hop
served 20.9k tokens across the network
hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:IQ3_MollamaIQ3_Mrung 2
bartowski · quantization of meta-llama/Llama-3.2-3B-Instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 71.97 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 15 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, faithful_tool_result, instruct.follow, lang.en, lang.hi, lang.kn, output.long, output.schema, research.synthesize, structured_output, tools, tools.call, tools.faithful, tools.loop_ends, tools_loop_terminated
served 4.0k tokens across the network
hf.co/unsloth/Phi-4-mini-instruct-GGUF:Q3_K_MollamaQ3_K_Mrung 2
unsloth · quantization of microsoft/Phi-4-mini-instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 66.57 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 6 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, lang.bn, lang.en, output.long, output.schema, structured_output
served 12.4k tokens across the network
hf.co/unsloth/Qwen3-4B-GGUF:IQ3_XXSollamaIQ3_XXSrung 2
unsloth · quantization of Qwen/Qwen3-4B
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 47.81 tok/s best · trained 40K · proved 32K ctx (recalled at every size asked) · 16 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.en, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools_loop_terminated
served 152.6k tokens across the network
hf.co/Qwen/Qwen3-4B-GGUF:Q5_K_MollamaQ5_K_Mrung 3
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 46.99 tok/s best · trained 40K · proved 32K ctx (recalled at every size asked) · 18 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.en, lang.te, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated
served 16.8k tokens across the network
hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_0ollamaQ4_0rung 3
unsloth · quantization of google/gemma-4-E2B-it
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 83.82 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 31 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.bn, lang.en, lang.gu, lang.hi, lang.kn, lang.ml, lang.mr, lang.pa, lang.ta, lang.te, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated, vision, vision.count, vision.multi_image, vision.read, vision.spatial
served 15.2k tokens across the network
Qwen/Qwen3-4BvllmBF16rung 3
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · proved 64K ctx (this machine could not serve at 128K) · 13 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, faithful_tool_result, lang.en, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.multi_hop
sarvamai/sarvam-1vllmBF16rung 3
sarvamai · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 8 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: lang.bn, lang.gu, lang.hi, lang.mr, lang.pa, output.long, output.schema, structured_output
hf.co/unsloth/Llama-3.2-3B-Instruct-GGUF:IQ1_SollamaIQ1_Srung 5
unsloth · quantization of meta-llama/Llama-3.2-3B-Instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 94.46 tok/s best · 1 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: lang.en
served 10.2k tokens across the network
hf.co/unsloth/Phi-4-mini-instruct-GGUF:Q2_KollamaQ2_Krung 5
unsloth · quantization of microsoft/Phi-4-mini-instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 68.61 tok/s best · 5 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, lang.en, output.long, output.schema, structured_output
served 27.9k tokens across the network
hf.co/unsloth/Qwen3-4B-GGUF:IQ1_SollamaIQ1_Srung 5
unsloth · quantization of Qwen/Qwen3-4B
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 61.33 tok/s best · on host
served 127.6k tokens across the network
hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF:Q3_K_LollamaQ3_K_Lrung 9
bartowski · quantization of meta-llama/Llama-3.2-3B-Instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 78.82 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 14 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, lang.en, lang.kn, output.long, output.schema, research.synthesize, structured_output, tools, tools.call, tools.faithful, tools.loop_ends, tools_loop_terminated
served 3.9k tokens across the network
hf.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF:Q3_K_LollamaQ3_K_Lrung 9
MaziyarPanahi · quantization of microsoft/Phi-4-mini-instruct
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 64.06 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 6 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.write, instruct.follow, lang.en, output.long, output.schema, structured_output
served 12.7k tokens across the network
hf.co/Qwen/Qwen3-4B-GGUF:Q5_0ollamaQ5_0rung 9
Qwen · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 47.96 tok/s best · trained 40K · proved 32K ctx (recalled at every size asked) · 18 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.en, lang.hi, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated
served 41.8k tokens across the network
hf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_XSollamaIQ4_XSrung 9
unsloth · quantization of google/gemma-4-E2B-it
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 79.01 tok/s best · trained 128K · proved 128K ctx (recalled at every size asked) · 31 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: code.reason, code.write, faithful_tool_result, instruct.follow, lang.bn, lang.en, lang.gu, lang.hi, lang.kn, lang.ml, lang.mr, lang.pa, lang.ta, lang.te, output.long, output.schema, reason.channel, research.synthesize, structured_output, thinking, tools, tools.call, tools.faithful, tools.loop_ends, tools.multi_hop, tools_loop_terminated, vision, vision.count, vision.multi_image, vision.read, vision.spatial
served 12.6k tokens across the network
kokorospeech-podrung 9
kokoro · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 3 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: audio, audio.out, lang.en
UsefulSensors/moonshine-basespeech-podrung 9
UsefulSensors · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 3 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: audio, audio.in, lang.en

Cohort 2 · Other vendor members · 1

UsefulSensors/moonshine-tinyspeech-podrung 9
UsefulSensors · vendor build
- Linux · x86-64 · NVIDIA · 32 GiB — 1 gram · 3 capabilities held · on host
held on Linux · x86-64 · NVIDIA · 32 GiB: audio, audio.in, lang.en

Every number is a signed measurement or a count of them. grams is how many distinct nodes contributed: a verified floor, not a complete account. Lineage (publisher, attribution, base, cohort, rung) is from the catalog and is not a measurement.

grams and classes are what THIS window heard from; coverage, refusals, cost and demand carry across windows, because a gram going to sleep does not unmake what it measured.

The season file → All seasons →