Markdown source

Dense Slides: Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning — Merve Noyan, Hugging Face

Source Video

Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning — Merve Noyan, Hugging Face

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

slide-001.jpg

Slide text:

Open-source

slide-002.jpg

Slide text:

Why does open-source matter?

Absolute control over models

Cost reduction in multipliers

Further customize, shrink models depending on your needs

Guaranteed privacy for the end-user, enable on-device/in-browser uses, data doesn’t go to another server

slide-003.jpg

Slide text:

Artificial Analysis Index

Artificial Analysis Intelligence Index by Open Weights / Proprietary 2 B 28 of 472 modeks

Artiflclal Analysls Intelligence Indcx v4.0 incorporates 10 evaluatlors: GDpval-AA, rr-Bench Telecom, Terminal- Barch Hard, SclCode, AA-LCR, AA-Omnlsclerca, IFBench, Humanity's Lest Exam, GPQA Dlamond, CritPt + Add modet from specifk provlder

Proprletary Open Welghts 8080808 A Artlflclw Autytts

5.0 49 49 Br

slide-004.jpg

Slide text:

Hugging Face Hub

Home for the open-source machine learning community: share & discover models, datasets, apps, connect with the community and more!

slide-005.jpg

Slide text:

models → agents, serve locally

Agentic LLMs (thinking + tool calling): gpt-oss, Gemma-4, Minimax M2.7, GLM-5, Nemotron3-Super

Agentic vision models (thinking + CUA): Qwen3.5

(Alibaba), Kimi-K2.5

mlx_ lm.generate --prompt *How tall ts Mt Everest?"

vllm serve Qwen/Qwen3-8B # then query with QpenAI Completion

llama-server -n model.gguf --port 8080

slide-006.jpg

Slide text:

compare open models

Datasets: @ScaleAl/SWE-bench_Pro like: 78 Follow @ ScaleA: 284

Benchmark Modalities: θ Text Formats: + parquet Size:<1K Libraries: & Datasets pandas. Polars. +1

t Dataset card: 田 Data Studio Flles and versions 1 K xot: O Community 0

. Leaderboard ioficisl Benchmrk ① Leam more Epedmentsl

Taslc SWE Bench Pro

HODEL SCORE

Φ Oza1-org/GLH-5.1:0 58.4

2 HiniMaxAI/niniMax-H2.5 T 55.4

3. Q coonshotai/KimI-K2.S n' 50.7

4 s Quen/Qnen3-Coder-Hext n: source 44.3

5 1 Qon/Qwan3-Coder-4888-A35B-Instruct n! tource 38.7

V Show all 14 modcls

slide-007.jpg

Slide text:

compare providers across models

slide-008.jpg

Slide text:

The Hub meets your agent

MCP Server

plug Hub in to your favorite LLM

HF CLI

search models, manage datasets & buckets, launch Spaces, run jobs.

Skills

empower your agent with Skills of HF ecosystem

Local Agents

Run full coding agents with llama.cpp, Pi & more

slide-010.jpg

Slide text:

Local coding agents Pi consumes llama.cpp Ilama-agent: agent loop baked into llama.cpp as binary 000 #add your local model $npm install -g @mariozechner/pi-coding-agent "providers":{ "llama-cpp":{ "baseUrl":"http://localhost:8080/v1", install pi configure

/build/bin/llama-agent -hf [HUB_ID] "models":[ "apikey":"none", "api":"openai-completions”,

(agent) pi API llama.cpp (server) "id":"Qwen3.5-122B-A10B-GGUF"

I your files, terminal, etc.

$pi #start pi start coding

slide-011.jpg

Slide text:

Local self-improving agents

Hermes Agent

self-improvement is baked in: the approach as a reusable across sessions after the task, the agent saves 'skill" and persists memory hermes chat --provider hf #use with inference providers HF_TOKEN=hf_... #usewith local served endpoint hermes config Set OPENAI BASE URL http://localhost:8080/v1 (llama.cpp & friends)

Providers or serve LM locally Integrated with Inference hermes config set OPENAI API KEY dummy name hermes config set LLM_MoDEL your-model-

#start chatting

hermes chat

slide-012.jpg

Slide text:

New! Hub hosts your agent traces

oher Talu Uonres Untute Uamsts: Datasets 21 a Fker by rume Ful-tet search,:, 1l Sort: Trendng

Hodalles 08 Audo: Documnt t + n6ttT - Lt9 a + or yrp eptpdn: el #: m badlogicganos/pi-aono + 0xSoro/pi-sossions +in· LpLta3dyxo· b%·±17. 5

D Geosptiul 9 Tet Tmeseries. a voto + Tnct· Updattdadp ato. a7- ±5a. 04 m badlogicganos/pi-diff-review jediscti/agent-traces-smival + Trca: Updaed about ghour ao · D T · A ll: 01

Sue(rour) 0 lhoestq/ag+nt-trac+s-+xaaple + Yaces: Updrid&day. sgo. D4· ±loe· 1 virhnx9e/vtcade-sassion± t o · rst + - alr tnou re noqr ptirpdn

Fo.k: + Trce · Updsttd sbout 1 boar sgo - a 1i · ± 1o · 1 noikapy/fxKobolds 食 tnoe- Updrted 6 dys ago: u 12· 主H cfahlgreni/egent-sessicns-list

Boe! = ov + prqut optimired parquet:. Bb tragetolder + Tnkes · Updned J dsy ago · D&· L39 mcfahlgron1/pi-diff-review + Tncgs + Updsted 1 day ago: a 15 + ±44 LaisEckart/appioraltostt-lava-sessions

》MTOH B soundfoldr: ■tit + Trces + Updated 1dsys hro - ct 42? + t(o1 m cfahlgrenl/pi-nono-fresh + Tract4 - Updsted 1 days sgo - B 1 - ± 3s M dongxx11e4/Basoline_featboach

Type Benchmark 1( ( Irxes x:± Trces·Updsted2 dnysato·02·上32 I davanstrien/pi-tracos M davanstrien/pi-traco-parsor-stssions ± Trx++Updt*d2dyt +go: aI+ ±2?

slide-013.jpg

Slide text:

Just upload your sessions from:

~/.claude/projects

~/.codex/sessions

~/.pi/agent/sessions

Nothing else needed!

slide-014.jpg

Slide text:

tip: find models supported by local apps

Hsin Isks Litrarirs Urpuges Lloemit (ther 0 Models 474 Tul et verch. y oun+ '+1 Sort, Trending.

Liu.coo G L Sudoe LnM ow Thinp O Hayr lop t9c 0 - ht1 F: otr kp t pvrpdn · 9st r + ju1-orl+lrul unsloth/geana-4-268-A48-it-GGuf D unsloth/gema-4-318-1t-GGUF P muItrl·to-Tit + d 1la - Updted sbout 18 hout ato · ± Ael. C H1

YSGLngQ Ualoh oiluionl+ Jbyfuion rUM om 口 xLxL t Dodar Hode tuaitar t Lemortde unsloth/gcana-4-E4B-it-GGUF D unsloth/goea-4-E2B-1t-GGUF B hugrTea-to-Yeat · d Se · Updstrd sbout 2l houn aro - ± y!lh. g1

tertmce Prerlders Select alt α unsloth/GLH-5.1-GGUF T Tta Gmrmtion - Tb. Updsted I dsy ao · a i1.y - O t unsloth/Q=en3.S-358-A3n-GGUF F tratrTento-Tent · 4 Jso - Updrted Hes· ±L4cy. O Ta

Croa Noic Cerebrns ○ ulX Hyperbolch TocetherA hmagt-Ten to-fer · d 2re · Updeted Ma 5· ±12k. O <s α untloth/Q=on3.S-27B-GGUF 0 uns1oth/Q=on3.5-98-GGUf B hmageler.to-Tet · s 9s - Updatrd Hst? · t l.1M. O 41?

t Osicloud A Endpoinau hF hitence AliD wrSped R Coher Fhtworis fetharlessA Q Zl D Rrpiratt Q Salcway d G d unsloth/Qaen3-Coder-Haxt-GGUF TetGanenAion · sce· Updxted Mr6· t 25sk · 556 Tet Goneraton - d 31B: Updned Jn 30 · t 13ek · 583 Q unsloth/Q=on3-Code1-30B-A3B-Inatrutt-GGUF

MiK unsloth/Qnon3.5-4B-GGUF kmwg+teat-to-Tor - 4B. Updstrd Har 2 - ± 5tdk - 19$ H hmaglanto-Tra · d 12ze · Updxrd r s -. 38yt. O 10 unsloth/Quon3.5-1228-A1e8-GGUF

hf.co/models → other → apps

slide-015.jpg

Slide text:

ggml-org/gemma-4-26B-A4B-it-GGUF uke 40 Folow ceml-org 168k

B GGUF: converatlonal

V Modelcard Ie Files and versions i& mt Community I:: Deploy * eUse this model]

Edit model card

gemma-4-26B-A4B-it-GGUF 67,075 Downlosds Iast monh

Recommended way to run this model:1lana-scrvor -hf ggal-org/gcnma-4-26B-A4B-it-GGUF GGUF ① Model size 25B params Architecture: gemma4 () Chat template

Then, access http://ocalhost:8080 Hardware compatibility P L4 (24 GB) x1:

4-bit Q4_K_M:16.8 GB

8-bit α Q8_0 126.9 G8:

16-bit β F16 150.5 GB

slide-016.jpg

Slide text:

lets any coding agent HF CLI Skill # Add CLI skill globally hf skills add --claude --global

search models, manage datasets, # 0r per-project hf skills add --claude

run jobs (and more!) launch Spaces # Works with other agents too hf skills add --codex hf skills add hf skills add --opencode --cursor

slide-019.jpg

Slide text:

Grest Do you hive sn ides of the costs foc runaing 16 GpUs in parstei tor thst long? Use @ Hugglng Face Jobt:

nq Uslng the latest rerifed HF Jobs pricing frorn our erlronmant, l4tsxl ks s1.8e/hour pe GPU. So 16x I4esx1 3 previous mosssges ) → LLM even does napkin calculates running cost for math for the instance and

Aep/169s noqy · +. $2s.te/hour tots the Job

Lyn yyap-its sinou oeioz hoqt ey pmoyt uru tded- ras'iz mm, out 'ehtd yyewlpuea sotl ueuro eu betn + -29.3h case: about s843 + 3+.th cso: sbout $565 16 jobs in paralle, to the rw GpU cost is about: ran you can use Buckets → these Jobs will be re-

For compsrson, 16x sieg-large: b ch+aotr par hour but slomt ottral: rd budget a bit abore rw compute ior shurd skew, slrtup overhead, retrles, trd possble Hub hlceupt + Practicel budgtl sbout s9eo-si,054 fast read-write with hf-mount that allows

+ $ltg-1srge.: s1.s0/bour. osch, so $24/bour. for 16

+ Estimated wall-clock: about 56.4 hours. + Total: sbout s1,353

Tthuo aodtoy> Aryoirw pue ostu, 4oq dq proyt 1' txhm xgt tsn:outt cn cirt uoepuauuoou Au os

Hidden Non-Slide Evidence

Classification audit: raw/sources/slide-ai-classification/dense/OV56RddyFuU/audit.json