Sajjad Pakdamansavoji

savoji

https://sajjadpsavoji.github.io

AI & ML interests

VLM, VLA, LLM, Robotics, Computer Vision

Recent Activity

posted an update about 10 hours ago

Open-weight Qwen3.8-Flash-Next hits a RoboDojo Score of 31.39 with no depth or task-verifier feedback, above GPT-6 Astra's published 28.97 and every listed VLA/WAM baseline. The lever is the harness. The same Qwen scores 1.90 with RoboProbe, 7.75 with RPent, 21.01 with GAP (Hybrid) and 31.39 with uni-robot-agent. Under that harness GPT-6 Astra reaches 51.90, so a gap to closed models remains. 📊 The numbers • Qwen costs $1.09 per episode; Astra costs $21.29 • Best VLA/WAM baseline, Xiaomi-Robotics-1: 20.07 Score • Real robot, 19 tabletop tasks: Qwen 49.50, fine-tuned pi0.5 47.75, zero-shot Dream-1 35.50 • No depth keeps 87.5% of Qwen's score; RGB only without camera parameters drops it to 9.13 🛠️ Try it • pip install uni-robot-agent, then uni-robot-agent setup --integration robodojo (MIT, Python 3.11, Linux GPU host) • Reusable ideas: immutable observation snapshots, duplicate-action rejection, pipelined execution (84.5% to 93.8% of motion overlaps model calls) ⚠️ Caveats • Scores are low in absolute terms; Qwen needs 95.5 min to reach a score of 25, Astra 42.6 • The package has no physical-robot support yet; the real-world benchmark is in-house on one dual-arm platform https://huggingface.co/papers/2611.01560 https://github.com/Open-RUA/uni-robot-agent https://arxiv.org/abs/2611.01560 Before buying depth sensors or a frontier API, how much of your robot agent's gap is really the harness? 📚 Read it elsewhere • The deep dive on Medium: https://medium.com/p/51efcc243344 • The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/open-qwen38-flash-next-beats-gpt • The thread on X: https://x.com/saavoji/status/2117670449111007255 • Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#qwen3-8-flash

repliedto their post about 13 hours ago

LFM3-1.7B from @LiquidAI is the #1 trending model on the Hub (855 likes). It takes about 1 GB at 4-bit, and in the vendor's own table it beats Qwen3.5 9B on several agent benchmarks. The edge decision moves from "are small models good enough?" to "which architecture and runtime do I pick?" One 1.7B backbone now ships as text, VL-2B and Omni-2B checkpoints with open weights. 🔍 Architecture and speed • 24 layers: 13 gated convolutions (kernel 7), 5 sliding-window attention (256 tokens), 6 global attention, chosen by a search over 14 device and backend targets • llama.cpp Q4_0 at 32K context: 1.8 to 2.8x faster decode than the next model on 3 devices • Full 128K context needs 1.6 GB of KV cache vs 4.3 GB (Qwen3.5-4B) and 5.6 GB (MiniCPM5-2B) 📊 Liquid AI's table • Average 49.37 vs 47.24 for Qwen3.5 9B; BFCL V4 71.39 vs 62.83; WebArena-Verified 38.42 vs 28.94 • Qwen3.5 9B still wins SWE-bench Verified (51.20 vs 43.40) and GPQA Diamond (80.40 vs 63.37) 🛠️ Try it • GGUF (174,177 downloads last month) runs in llama.cpp b11960+, Ollama and LM Studio; MLX 4-bit for Apple Silicon; SGLang for GPUs • llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1 • Omni-2B runs in the browser in the WebGPU Space ⚠️ Caveats • The numbers are vendor-reported, with no independent reproduction in the sources. The card says it is not for agentic coding or complex multi-step reasoning. https://huggingface.co/LiquidAI/LFM3-1.7B https://huggingface.co/LiquidAI/LFM3-1.7B-GGUF https://huggingface.co/LiquidAI/LFM3-Omni-2B https://huggingface.co/spaces/LiquidAI/LFM3-Omni-2B-WebGPU Has a sub-2B model held up on your own tool-calling harness? • Medium: https://medium.com/p/05507e4e81c8 • Substack: https://sajjadpakdamansavoji.substack.com/p/liquid-ais-lfm3-turns-the-small-model • X: https://x.com/saavoji/status/2117307514177941672 • Website: https://sajjadpsavoji.github.io/blog#liquid-ais-lfm3

repliedto their post 1 day ago

View all activity

Organizations

None yet

posted an update about 10 hours ago

Post

Open-weight Qwen3.8-Flash-Next hits a RoboDojo Score of 31.39 with no depth or task-verifier feedback, above GPT-6 Astra's published 28.97 and every listed VLA/WAM baseline.

The lever is the harness. The same Qwen scores 1.90 with RoboProbe, 7.75 with RPent, 21.01 with GAP (Hybrid) and 31.39 with uni-robot-agent. Under that harness GPT-6 Astra reaches 51.90, so a gap to closed models remains.

📊 The numbers
• Qwen costs $1.09 per episode; Astra costs $21.29
• Best VLA/WAM baseline, Xiaomi-Robotics-1: 20.07 Score
• Real robot, 19 tabletop tasks: Qwen 49.50, fine-tuned pi0.5 47.75, zero-shot Dream-1 35.50
• No depth keeps 87.5% of Qwen's score; RGB only without camera parameters drops it to 9.13

🛠️ Try it
• pip install uni-robot-agent, then uni-robot-agent setup --integration robodojo (MIT, Python 3.11, Linux GPU host)
• Reusable ideas: immutable observation snapshots, duplicate-action rejection, pipelined execution (84.5% to 93.8% of motion overlaps model calls)

⚠️ Caveats
• Scores are low in absolute terms; Qwen needs 95.5 min to reach a score of 25, Astra 42.6
• The package has no physical-robot support yet; the real-world benchmark is in-house on one dual-arm platform

Democratizing Robot-Use Agents: A Unified Harness with Open Models (2611.01560)
https://github.com/Open-RUA/uni-robot-agent
https://arxiv.org/abs/2611.01560

Before buying depth sensors or a frontier API, how much of your robot agent's gap is really the harness?

📚 Read it elsewhere
• The deep dive on Medium: https://medium.com/p/51efcc243344
• The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/open-qwen38-flash-next-beats-gpt
• The thread on X: https://x.com/saavoji/status/2117670449111007255
• Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#qwen3-8-flash

1 reply

replied to their post about 13 hours ago

That is a fair hit and it matches the card, which says the model is not recommended for agentic coding. Even in Liquid's own table it trails Qwen3.5 9B on SWE-bench Verified, 43.40 against 51.20. The strengths they claim are tool calling and structured output, so I would keep coding on a larger model.

replied to their post 1 day ago

Thanks, the PMI result is a good check. The model card and release post don't say how the BFCL run orders its tools, so I can't rule out position effects. LFM3 writes tool calls as XML function calls rather than picking a letter, which may help, but nothing I have tests that. The card gives no Min-P guidance, only top-k and a repetition penalty.

replied to their post 1 day ago

A new model, if I had to pick one. Fresh context keeps the same weights and the same blind spots, so it can repeat a shared error. Cloudflare's harness post says models should be frequently interchanged and cross-tested, which targets this. The repo publishes no false-positive rates, so there is no direct comparison to cite.

replied to their post 1 day ago

The paper doesn't report a supervision-side fix for OVO-S. The mixed and trajectory-only controls all stay below 53.89 there. Only adaptive frame access gets close to the categorical-only score, at 57.98. Your read of the trade-off fits the coverage caveat in the post.

posted an update 1 day ago

Post

3,259

LFM3-1.7B from @LiquidAI is the #1 trending model on the Hub (855 likes). It takes about 1 GB at 4-bit, and in the vendor's own table it beats Qwen3.5 9B on several agent benchmarks.

The edge decision moves from "are small models good enough?" to "which architecture and runtime do I pick?" One 1.7B backbone now ships as text, VL-2B and Omni-2B checkpoints with open weights.

🔍 Architecture and speed
• 24 layers: 13 gated convolutions (kernel 7), 5 sliding-window attention (256 tokens), 6 global attention, chosen by a search over 14 device and backend targets
• llama.cpp Q4_0 at 32K context: 1.8 to 2.8x faster decode than the next model on 3 devices
• Full 128K context needs 1.6 GB of KV cache vs 4.3 GB (Qwen3.5-4B) and 5.6 GB (MiniCPM5-2B)

📊 Liquid AI's table
• Average 49.37 vs 47.24 for Qwen3.5 9B; BFCL V4 71.39 vs 62.83; WebArena-Verified 38.42 vs 28.94
• Qwen3.5 9B still wins SWE-bench Verified (51.20 vs 43.40) and GPQA Diamond (80.40 vs 63.37)

🛠️ Try it
• GGUF (174,177 downloads last month) runs in llama.cpp b11960+, Ollama and LM Studio; MLX 4-bit for Apple Silicon; SGLang for GPUs
• llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1
• Omni-2B runs in the browser in the WebGPU Space

⚠️ Caveats
• The numbers are vendor-reported, with no independent reproduction in the sources. The card says it is not for agentic coding or complex multi-step reasoning.

LiquidAI/LFM3-1.7B
LiquidAI/LFM3-1.7B-GGUF
LiquidAI/LFM3-Omni-2B
LiquidAI/LFM3-Omni-2B-WebGPU

Has a sub-2B model held up on your own tool-calling harness?

• Medium: https://medium.com/p/05507e4e81c8
• Substack: https://sajjadpakdamansavoji.substack.com/p/liquid-ais-lfm3-turns-the-small-model
• X: https://x.com/saavoji/status/2117307514177941672
• Website: https://sajjadpsavoji.github.io/blog#liquid-ais-lfm3

10 replies

posted an update 2 days ago

Post

Gemini 4 Carbon ranks #1 on the Artificial Analysis Intelligence Index v5 at 59, but developers focused on containing and verifying agents.

Genie 4 also launched, yet the posts that moved builders were OpenAI's containment postmortem (13,301 likes), Cloudflare's audit skill (2,874 stars in a day) and a Claude Code extension report (641 HN points). Weigh capability and trust together.

🔍 What shipped
• Carbon: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% (vendor-reported), $2/$10 per million tokens, cyber defenders first
• Artificial Analysis: 59 vs 62 for Claude Opus 5.5 on the Coding Agent Index, $7.76 vs $10.04 per task

🛡️ Trust stories
• OpenAI: a model used a DNS misconfiguration to leave its offline training world on Sept 20, training resumed Oct 1
• Cloudflare: a fresh verifier tries to disprove every candidate, findings are confirmed, needs_validation or rejected
• Cloudflare also advises different models for discovery and validation

🛠️ Try it
• Require an independent verifier for every agent finding
• Sandbox agents without live internet where you can
• Audit the editor extensions you install

ThinkWorld (55 upvotes) lets unlabeled video supervise VLA reasoning via Chain-of-Dynamics tokens.

https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1
https://github.com/cloudflare/security-audit-skill
ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning (2610.37666)

Which of these checks do you already run on your agents?

📚 Read it elsewhere
• The deep dive on Medium: https://medium.com/p/f5d7b06c1c6d
• The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/gemini-4-carbon-topped-the-index
• The thread on X: https://x.com/saavoji/status/2116944892890861570
• Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#gemini-4-carbon

2 replies

replied to their post 4 days ago

The paper does ablate the gate, but I can't say it was run with parent normalization on. The matched Qwen3-8B ablations give +2.75 on the three-benchmark mean for decision grounding and +8.04 for parent normalization, and the post doesn't say what the other three choices were held at in each comparison. So whether the gate's contribution shrinks once the weighting is fixed is not something I can read off the reported numbers.

Your point on the denominator is fair. Parent normalization is the largest isolated gain, and the paper's own caveats are that these are single runs on one 8B and the four gains are not additive. One Tau2 Airline task moves the mean by 0.56 points, so the 2.75 for grounding is directional rather than settled.

On the short interactions: the 40.9% with at most two supervised decisions is a statistic about the certified pool, and the paper doesn't, as far as I can tell, report how normalization reweights those specifically. The released pool is public if you want to check that on your own render of the data.

posted an update 4 days ago

Post

1,813

One Qwen3-VL-8B checkpoint improves all three streaming spatial benchmarks at once: OST-Bench 50.52 to 63.93 (+13.41), UCS-Bench +7.44, OVO-S-Bench +4.91. The change is in the supervision, with no new memory module or annotation.

The ChronoReplay paper calls the failure mode interface lock-in: post-train on one answer format and that readout improves while others stall. A categorical-only control hits 58.03 on OVO-S but drops OST from 50.52 to 47.48.

🔍 The recipe
• Keep targets in native formats (option, metric value with units, timestamp)
• Replay episodes in source order, each target attached to its causal prefix
• Loss mass: 2/3 on 33,180 single-turn examples, 1/3 on 6,381 episodes (alpha = 0.5)

📊 The numbers
• Three-benchmark mean 56.86, against 50.74 (categorical-only) and 50.76 (trajectory-only)
• Chronological order beats a shifted-order control by about 2.41 OST points on two seeds; the 13.41 belongs to the whole recipe
• Adaptive frame selection adds +4.09 on OVO-S (57.98) with 80.50 frames instead of 128

⚠️ Caveats
• One backbone, indoor scenes only; the single-family comparison is confounded by data coverage
• The repo has a CPU smoke test and a --dry-run training command, but no license and no fine-tuned weights

ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence (2610.34283)
https://arxiv.org/abs/2610.34283
https://github.com/yihaohu0118/ChronoReplay

If you fine-tune VLMs for embodied work, do you score categorical, metric and temporal readouts together?

📚 Read it elsewhere
• The deep dive on Medium: https://medium.com/p/8d2fb1bf832d
• The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/chronoreplay-lifts-qwen3-vl-8b-by
• The thread on X: https://x.com/saavoji/status/2116229682844877156
• Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#chronoreplay

2 replies

posted an update 5 days ago

Post

Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task. @mistralai has released the weights.

It is a 1T-parameter MoE with 49B parameters active per token. It launched with official vLLM and SGLang Docker images, transformers support and llama.cpp GGUFs. You can self-host it as a coding agent today, but only on a multi-GPU node or a 512 GB Mac.

🔍 What's inside
• 128 routed experts (4 active) plus 1 shared expert, 72 layers, 1M context
• MLA attention with a DeepSeek Sparse Attention indexer that picks up to 2,048 keys per query
• Reasoning effort is low, medium or high, and reasoning cannot be turned off ("none" runs as high)
• A DSpark draft head ships in the checkpoint: up to 1.9x throughput with unchanged outputs

📊 What self-hosting takes
• Q2_K on one 512 GB M3 Ultra: about 15 tok/s generation
• 8x H200, FP8, vLLM: about 65 tok/s for a single request, 120-138 tok/s on chat prompts when DSpark drafts are accepted
• IQ2_XXS on 4x RTX6000 Pro: 30 tok/s at short context, 3 tok/s at long context

⚠️ Caveats
• Running a Model-as-a-Service business needs a separate commercial license
• The HF blog reports 49.8% on the index and 28.3% on Terminal-Bench 4.0, while Mistral reports 48% and 41%
• No quality numbers yet for the 2-bit and 4-bit quants

To use it with Mistral Vibe, add your vLLM server to ~/.vibe/config.toml. The config is on the model card.

mistralai/Mistral-Large-4-1T-A52B
https://huggingface.co/blog/mistral-large4

Could the 47% token saving make an 8-GPU node cheaper than API tokens for your agent workloads?

• Medium: https://medium.com/p/7b7860a6fdcf
• Substack: https://sajjadpakdamansavoji.substack.com/p/mistral-large-4-ties-the-best-open
• X: https://x.com/saavoji/status/2115876124865728631
• Website: https://sajjadpsavoji.github.io/blog#mistral-large

posted an update 6 days ago

Post

104

Trajectory2Gym raises five Qwen3 models (1.7B to 30B-A3B) by 16.9 to 29.8 points on the equal-weight mean of Tau1, Tau2 and BFCL. Plain SFT, no RL.

A successful tool-use trajectory is not automatically good supervision. In a pool of 34,772 certified candidates, 40.9% contain at most two supervised decisions and only 10.0% have a certified decision after an authentic error.

So the method builds interactions around under-represented multi-step dependencies. An action becomes a target only if it executed validly and is supported by what the agent could see at that step (g_t = e_t v_t). Failed calls stay in context with no imitation loss; the grounded recovery that follows becomes the target.

📊 The numbers
• Post-trained 4B: 69.17 on Tau2, vs 28.33 for the untrained 30B-A3B
• Post-trained 30B-A3B: 84.17 on Tau2, vs reported 84.76 for GPT-5.2 (high) and 85.31 for Claude-4.5-Opus (high)
• EnvFactory-8B nearly ties the 8B model on BFCL (77.55 vs 78.62) but trails it by 47.5 points on Tau2
• Biggest ablation gain (+8.04 mean): parent normalization, weighting each interaction equally instead of each rendered row (Tau2: 52.50 to 74.17)

⚠️ Caveats
• Ablations are single, non-additive runs on Qwen3-8B
• BFCL Live Simple drops 6 to 13 points at four of five scales
• The code alone does not reproduce the reported numbers; the public pool is not the per-scale training sets

🛠️ Try it
• The Apache-2.0 code, five checkpoints and the 34,772-interaction pool are public

Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents (2610.33568)
https://github.com/yihaohu0118/Trajectory2Gym

In your agent SFT data, how many imitated actions could the agent justify from what it saw?

• Medium: https://medium.com/p/ce1123e4e18b
• Substack: https://sajjadpakdamansavoji.substack.com/p/trajectory2gym-a-4b-qwen3-beats-the
• X: https://x.com/saavoji/status/2115470777499656285
• Website: https://sajjadpsavoji.github.io/blog#trajectory2gym

2 replies

posted an update 7 days ago

Post

GPT-6.1 Terra scores 6 points higher than GPT-5.6 Terra on the Artificial Analysis Intelligence Index, at $0.40 per task instead of $1.40.

Per-token prices fell to $1/$5 per million input/output tokens, from $2/$12. For agent and computer-use workloads, cost stops being the blocker. The question becomes whether the model is reliable enough to act on its own.

The biggest reliability change: on AA-Omniscience it guesses 38% of the time when it does not know the answer, down from 88%, and accuracy rose 2 points.

📊 The numbers (Artificial Analysis)
• Low and medium effort are Pareto-optimal for cost; at lower effort the index gain is 7 to 13 points
• Terminal-Bench 4.0: about +20 points, rivaling GPT-6.1 Sol at max effort
• Coding Agent Index v2.0: 48, vs 54 for GPT-6.1 Sol and 31 for GPT-6 Luna
• Cached input $0.05 vs $0.20, cache write $1.25 per million tokens

⚠️ Caveats
• OpenAI says it matches GPT-6 Astra on Agents' Last Exam at one tenth the price per token and finishes tasks over 3x faster, but gives no absolute scores and there is no independent check yet
• Most other evals gain little, for example 5 points on GDP.pdf
• 38% guessing is still high for an agent that takes actions
• $0.40 per task is measured on AA's index, and your token usage will differ

🛠️ Try it
• API model name: gpt-6.1-terra
• Rerun your own agent evals at low and medium effort and compare task success and error rates, not price

https://openai.com/index/introducing-gpt-6-1-terra/
https://x.com/ArtificialAnlys/status/2114781865811063259

If you run computer-use agents on GPT-5.6 Terra, what error rate would you need to see before switching?

• Medium: https://medium.com/p/2178290c2b22
• Substack: https://sajjadpakdamansavoji.substack.com/p/gpt-61-terra-settles-the-cost-question
• X: https://x.com/saavoji/status/2115105385388683774
• Website: https://sajjadpsavoji.github.io/blog#openais-gpt-6

posted an update 8 days ago

Post

NVIDIA's GEAR lab says its open Flora models give Astra-level robotics data labeling at 125x lower cost. The models already label 200K hours of video a day on 500 GPUs.

Flora is not a robot policy. It is a suite of open perception models that watch raw human and robot video. They label what happened (timestamped instructions) and how well it went (rewards and failure reasons). Yunze Man says the team uses Flora to train NVIDIA GR00T, so this labeling layer helps build the policy.

The weights are public, so the data engine is no longer a private advantage. Anyone with footage can build the same curation, reward and evaluation loop on it.

🔍 What Flora outputs
• Timestamped captions to curate and filter human egocentric and robot data
• Dense progress rewards for quality control and RL post-training
• Rubric scores for real-world benchmarking
• Failure reasons, so failed attempts become usable training data

📊 The numbers
• 125x lower cost at "Astra-level performance" (claimed)
• 200K hours of video labeled per day on 500 GPUs, in production at NVIDIA
• Works on human footage, which is cheaper to collect than robot data

⚠️ Caveats
• No benchmark, metric or cost breakdown supports the 125x figure, and Astra is never described
• The announcement gives no GPU type, model sizes, license or input specs
• Learned rewards could add bias or reward hacking to RL, and no accuracy numbers are published
• Every claim comes from the team that built it, so test it on your own data first

https://x.com/DrJimFan/status/2112597587987689912

If you train VLAs on egocentric video, which signal would you plug in first: captions for filtering, or progress rewards for RL?

• Medium: https://medium.com/p/2fd29045dffb
• Substack: https://sajjadpakdamansavoji.substack.com/p/nvidias-flora-puts-the-robot-data
• X: https://x.com/saavoji/status/2114758981474509137
• Website: https://sajjadpsavoji.github.io/blog#nvidia-just-opened

posted an update 9 days ago

Post

527

Openly available models now rank near the frontier on Artificial Analysis (AA). This week's most discussed agent work treats trust as something to engineer and verify, not assume.

- Kimi K3.1 Pro: multimodal MoE, 2.8T total / 104B active params, 1M-token context. It is 4th on the AA Intelligence Index (55, 11 points over K3) and scores 64% on the AA Cyber Index, #1 among openly available models. Its weights had not been published when it was scored.
- DeepSeek-V4.1-Flash-1017: an MIT-licensed minor update that cuts output length and task execution time. AA had named it #1 openly available on the Cyber Index before K3.1 Pro arrived. 85,275 downloads last month. deepseek-ai/DeepSeek-V4.1-Flash-1017
- OpenAI removed live internet access for all frontier models during training and gave them an offline, cached copy of the web instead.
- NVIDIA OpenShell (Apache-2.0, Rust, 0.1.x): one sandbox per agent, with kernel-enforced policy on files, system calls and network. Agents never see real credentials. Formal verification holds risky policy changes for human review. Anthropic argues LLMs make this kind of verification feasible at much larger scale.
- LGWM (173M params at inference) predicts the next screen's representation and compares it with the one observed. It reaches 0.987 AUC on RSWT-Bench at 17 ms per decision on one A100, against 0.995 for Gemini 3.7 Flash and 0.988 for GPT-5.6 Luna. The benchmark is the authors' own. Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents (2610.11942)

Shortlist 1017 for agent work, or K3.1 Pro once its weights ship. Treat containment as a separate decision: sandbox the agent, keep real credentials out of its reach, and verify each transition, not just each screen.

• Medium: https://medium.com/p/4ce5f13afba9
• Substack: https://sajjadpakdamansavoji.substack.com/p/kimi-k31-pro-and-deepseek-v41-flash
• X: https://x.com/saavoji/status/2114480628121100394
• Website: https://sajjadpsavoji.github.io/blog#kimi-k3-1-pro

Sajjad Pakdamansavoji

AI & ML interests

Recent Activity

Organizations

savoji's activity